CONNECTING THE DOTS

Alexander R. Backus

The mechanisms of associative memory in hippocampus and neocortex

CONNECTING THEDOTSThe mechanisms of associative memory

in hippocampus and neocortex

Alexander R. Backus

The work described in this thesis was carried out at the Donders Institute for Brain,Cognition and Behaviour, RadboudUniversity Nijmegen, with financial support fromthe European Research Council (ERC-StG 261177) and the Netherlands Organisationfor Scientific Research (NWO-Vidi 452-12-009) awarded to Christian Doeller.

ISBN/EAN 978-94-6284-119-2

Printed by proefschriftmaken.nl

Copyright © Alexander R. Backus, 2017

CONNECTING THEDOTSThe mechanisms of associative memory

in hippocampus and neocortex

Proefschrift

ter verkrijging van de graad van doctoraan de Radboud Universiteit Nijmegen

op gezag van de rector magnificus prof. dr. J.H.J.M. van Krieken,volgens besluit van het college van decanen

in het openbaar te verdedigen op vrijdag 22 september 2017om 10:30 uur precies

door

Alexander Rudolph Backusgeboren op 7 maart 1986

te Zeist

Promotor

Prof. dr. D.G. Norris

Copromotor

Prof. dr. C.F.A. Döller

Manuscriptcommissie

Prof. dr. G.S.E. FernándezProf. dr. N. Axmacher (Ruhr-Universität Bochum, Duitsland)Dr. A. Takashima

1.

2.

3.

4.

5.

6.

7.

8.

Memory can be viewed as a network of associations. (this thesis)

Remembering is subserved by a network of interacting brain regions, including the hippocampus and medial prefrontal cortex. (this thesis)

The hippocampus acts as a convergence zone of mnemonic information. (this thesis)

Associative learning involves a neural pattern change, where associated representations become more alike. (this thesis)

Theta oscillations mediate the interplay between the hippocampus and medial prefrontal cortex to facilitate the integration of disparate memories. (this thesis)

Whenever observed brain activity is bilateral, it must mean something. (Sander E. Bosch)

A bold stripe shirt calls for solid colored or discreetly patterned suits and ties. (Patrick Bateman)

The most important question in science is your question. (György Buzsáki)

Propositionsaccompanying the dissertation

CONNECTING THE DOTSThe mechanisms of associative memory in

hippocampus and neocortex

byAlexander R. Backus

Foreword

In his smashing debut, Alex Backus takes us on a scientific tour-de-forcethrough the realms of associative memory research. Though each chap-ter is information-dense and action-packed, Backus charms the readerwith succinct yet sprightly prose. Many will devour this deliciously play-ful but mysterious little thesis from cover to cover.

Sander E. Bosch, PhD

Locally-renowned neuroscientist and karaoke singer

v

La mémoire, c’est ce qui reste quand on a tout oublié.

Maître François-Xavier Boudringhin2 mai 1926 - 8 juin 2015

Grand-oncle de ma famille

Preface

”Science is like free-market capitalism” ¹. Onemight wonder what thesewords have to do with this dissertation. In all honesty, my academicjourney has been significantly affected by the brisk pace, volatile re-search trends, sentiment and opportunities to impact the cognitive neu-roscience of memory field, and has thus culminated in the medley ofexperimental chapters presented here. Needless to say, the topic of hu-man associative memory extends beyond the limited collection of as-pects touched upon in this work. Unraveling the neural mechanisms ofmemory is comparable to solving a complex puzzle with many dimen-sions, ambiguous pieces and occasional overlap with other unknownpuzzles. It is a very important puzzle, fundamental to human nature.This thesis embodies my attempt to add small but significant pieces todifferent sides of this puzzle. These are the pieces that help to connectthe dots: how the brain connectsmemories and howmemories connectthe brain.

¹Christian Doeller, personal communication on April 1st, 2012

ix

Contents

Prologue 1

1 General introduction 31.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.2 Memory concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.3 Memory representations . . . . . . . . . . . . . . . . . . . . . . . . 81.4 The hippocampus . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101.5 Memory integration. . . . . . . . . . . . . . . . . . . . . . . . . . . 171.6 Theta oscillations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201.7 Thesis outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2 Mnemonic convergence in the human hippocampus 272.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352.4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372.5 Supplemental Information . . . . . . . . . . . . . . . . . . . . . . . 45

3 Increased hippocampal neural pattern similarity of newly associ-ated stimuli 513.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 563.4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593.5 Supplemental Information . . . . . . . . . . . . . . . . . . . . . . . 65

Mesologue 67

4 Hippocampal-Prefrontal Theta Oscillations Support Memory Inte-gration 694.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 714.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 774.4 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814.5 Supplemental Information . . . . . . . . . . . . . . . . . . . . . . . 84

xi

xii

5 General discussion 955.1 Neuroscientific insights . . . . . . . . . . . . . . . . . . . . . . . . . 965.2 Methodological advancements . . . . . . . . . . . . . . . . . . . . . 1025.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1045.4 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

Epilogue 109

References 111

Nederlandse samenvatting 125

List of publications 129

Acknowledgements 131

CurriculumVitae 135

Donders Graduate School for Cognitive Neuroscience 137

Prologue

November 7th, 2013, San Diego, California, USA.

Peanut butter soup with smoked duck and mashed squash. New YorkMatinee called it a ”playful but mysterious little dish”. We followed-upwith red snapper with violets and pine nuts, I recall. A pleasant Novem-ber breeze swept across the San Diago bay while we were having din-ner aboard the Admiral Hornblower, a flagship vessel modeled after atraditional steamboat. The sumptuous waterborne dinner formed theconcluding stage of the Society for Neuroscience international con-ference satellite event, against the scenic backdrop of the city andbay area brought to us from the Admiral’s bow deck. When the din-ner commenced, I managed to secure a table seat next to the world’smost renowned hippocampus electrophysiologist, with some help ofmy wingman Toby. Notable other table party members were the ever-sociable friend-of-the-lab Raphael and a random Canadian-Egyptianneurosurgeon. The relaxed and informal atmosphere allowed for vi-brant discussions and mandatory microlectures by the grandmasterhimself. Various topicswere discussed; ample questionswere answered.But one question remained unanswered: how, in the name of Science,am I able to remember all of this?

1

1General introduction

3

1

4

1.1. IntroductionHumans have the amazing ability to encode, retain and retrieve massive amountsof information about past events (Tulving et al., 1972). Our brains are biological big-data¹ information-retrieval systems, each able to store and process approximately2.5 petabytes worth of data in the form of memories (Reber, 2010). To give an im-pression: this amount of data is equivalent to millions of hours of television, or theannual data harvest of one-hundred Donders-sized neuroimaging-research centers.In general, we are able to access the majority of our memories at will, and extractknowledge from it to interact with the environment in a deliberate way. Ultimately,the unique amalgam of memories, stored in our individual brains, lies at the heart ofour personality (Doeller, 2015). It defines us as human beings, for the mind is the in-evitable product of its nature (i.e. the genetic blueprint) and nurture (i.e. experience-dependent contributions in the form of memories). Needless to say, memory isubiquitous in our daily lives. For these reasons, elucidating the brain mechanismssupporting memory and further our understanding of this important cognitive func-tion is a worthwhile scientific exercise and constitutes the general overarching goalof this dissertation.

More specifically, the work presented in this thesis focuses on the key motifof memory: associations (Plato, Theaetetus). Associations between mnemonic ele-ments form the building blocks of memories about past events (Tulving et al., 1972).For instance, let us consider an everyday-life event that has taken place at a certaintime and location, involving a specific group of people. The associations betweenthese features characterize the event (i.e. the what, where and when). The brainmechanisms that support the formation, storage and retrieval of these associationsare part of a complex puzzle of the neural coding principles and intercommunicat-ing brain regions. How are mnemonic associations represented in the brain? Andhow do brain regions interact to support the encoding and retrieval of these as-sociations? This thesis revolves around these two key questions. From one pointof view, I investigated the connected representations in the brain that form memo-ries, while in parallel, I studied the connectivity between brain regions that underpinthese memory functions. But before getting to the answers to these questions, thetopic of associative memory is introduced in more detail in the remainder of thisintroductory chapter.

The forthcoming section is devoted to delineating the concepts of memory, be-fore I discuss the building blocks of memory: conjunctive representations of asso-ciations between mnemonic elements. Next, I review the key brain structures thatharbor these representations and explain how cognitive neuroscientists can experi-mentally probe them. I will then shift focus to the connectivity profile of the brain

¹By present-day standards c.f. Moore’s Law.

1

5

regions involved in memory and touch upon the methods used to analyze commu-nication networks. Next, I move on to the special case of memory integration, whentwo or more separate episodic memories are linked and recombined to form a new,extended memory network. In addition, I will highlight the functional interactionsbetween brain regions that support memory integration, and the potential role ofbrain oscillations in facilitating interregional communication. Throughout this intro-duction these topics are linked to the experimental chapters. Finally, I summarizethe specific research questions and experimental chapters of this thesis.

1.2.Memory conceptsWhat is memory? Prior to embarking on neuroscientific inquiry of cognitive phe-nomena, we must first attempt to delineate the core concepts under study. To thisend, I will evaluate two broad definitions of the general concept of memory, beforeelaborating on the specific aspect of memory that is studied in this thesis: associa-tions. One possible definition of memory is the retention, over time, of experience-dependent internal representations (Roediger et al., 2007). This definition empha-sizes the role of time and representations. Internal representations encode somesort of information, data, about the external world. Although quite specific on thesubject of study (i.e. experience-dependent internal representations), according tothis definition, a book also has memory with written words as internal represen-tations. Arguably, this is not the type of memory we aim to investigate, since weare interested in neural mechanisms underlying memory. Moreover, the definitiondoes not address the behavioral consequence of memory, since a representationcould hypothetically be retained over time, without ever having been retrieved toguide behavior. Alternatively, from a systems perspective, we could define memoryas the neurocognitive capacity to encode, store and retrieve information (Roedigeret al., 2007). This definition is broad enough to capture the folk-psychological ideaof memory, referred to during everyday discourse. Although the definition invokesthree arbitrarily separated stages of memory (i.e. encoding, storage and retrieval),it might however not be specific enough to isolate an aspect of function of mem-ory for scientific investigation. During encoding, information is stored as a lastinginternal representation. Storage involves the consolidation of memories to renderthem stable and resistant to interference (Frankland and Bontempi, 2005). Duringretrieval, this stored representation is accessed and reactivated to guide behavior(Tulving et al., 1983). Problematically, during each of the stages many other cognitiveprocesses² are likely involved, such as attention, emotion, perception and decision-making (Peelen and Kastner, 2014). In sum, due to its ubiquitous nature, memory

²Equally ill-defined, without exceptions.

1

6

may be regarded as a fuzzy concept, lacking a precisely circumscribed meaning³ .Nevertheless, although memory may be difficult to define, scientific inquiry basedon the broader definitions can still be fruitful. As long as we are aware of their lim-itations, we can work with both aforementioned definitions of memory. Next, wewill see how different types of memory can be subdivided.

One of the most influential taxonomies of memory proposes a distinction be-tween declarative memory and procedural memory (Squire, 1987). Procedural mem-ories are implicit and can not be articulated. For example, a memory of how to ridea bike is automatically retrieved without conscious awareness. Conversely, declara-tive memories are consciously accessible and introspectively reportable. Accordingto this taxonomy, declarative memory can be further subdivided into episodic andsemantic memory. Episodic memories have coordinates in space and time. For in-stance, the memory of a specific dinner cruise in San Diego is episodic. Semanticmemories have a more factual character. For instance, conceptual knowledge ofmathematics is independent of a spatiotemporal context. However, importantly, inthe real world, semanticmemories emerge from repeated episodic exposure (Tulvinget al., 1972). By extracting commonalities across a collection of disparate episodes,we are able to generalize factual information. For example, consider the fact thatthe city of Nijmegen is lying astride the Waal river. You might have made this obser-vation while viewing Google Maps, learned the information from your high schoolteacher, or read the sign ”Als de waal in het zicht is stroomt de verbeelding” when youcrossed the rail bridge accessing the Nijmegen train station. Accordingly, the bound-ary between episodic and semantic memory is arguably less clear-cut than proposed.Overlap between theoretically distinct types of memory makes it difficult to isolatetheir cognitive components and thus hinders neuroscientific inquiry.

The work presented in this thesis focuses on basic associative memory, definedas a person’s ability to explicitly learn and remember the relationship between pre-viously unrelated items (see Box 1). These memories are declarative, and form thebuilding blocks of episodic and ultimately semantic memories. Here, I viewmemoryas a chain (or network) of associations (or links) between hierarchical mnemonic el-ements (or nodes): a memory space (Boring, 1950; Tolman, 1948; Eichenbaum et al.,1999; O’Keefe and Nadel, 1978) (Figure 1.1). This memory space, with its networkstructure of associations, is used as working definition of memory throughout thisthesis.

Episodic memories comprise associations between events and their spatiotem-poral context (Tulving et al., 1972). In addition, one episodic memory can be asso-ciated with a second episodic memory, through partially overlapping features orcontext, forming links between separate memories (Preston and Eichenbaum, 2013;

³I owe this invaluable insight to the DoellerLab reading class.

1

7

FEATURES EPISODES CONCEPTS

time

food

people

location

flight

dinner

conference

lobby

read

new york

brain

scientist

Figure 1.1: Network structure of memory.Left: example network graph of associations characterizing a single episode, in this case a dinner cruisewith a scientist. The episode is characterized by features (e.g. people and food), which are linked tounique spatiotemporal coordinates (time and location). Each feature (a mnemonic element) is repre-sented as a node, whereas an association between elements results in an edge, with different strengths(thicknesses). The memory space is hierarchically organized, representing associative information atmultiple scales (middle and right graph). The episode is connected to other episodes via overlappingfeatures, such as the people or the location (middle). In turn, regularities across these connectedepisodes comprise the more abstract knowledge representation of the scientist (concept level) withits own interrelated concepts (right). Note that boundaries are simplified for display purposes: anarbitrary number of levels can exist below, above and inbetween the three levels sketched here.

Kumaran and McClelland, 2012). These interlinked memories, together ultimatelygeneralize to context-independent semantic memories and give rise to our knowl-edge base (Kumaran et al., 2009; Tenenbaum et al., 2011). Therefore, by studyingbasic associative memory, the goal of this thesis is to improve our fundamental un-derstanding of the whole breadth of cognitive capabilities related to memory. Nowthat we are familiar the type of memory under study, I will review how these asso-ciative memories might be represented in the brain.

Box 1: The paired-associate learning paradigm

The principal paradigm to study associative memory is paired-associate learning(Calkins, 1894; Hulse and Deese, 1967). In a paired-associate learning task, partici-pants are asked to learn the pairing between two previously unrelated items. Afterlearning, the person is prompted with one of the items (the cue) and has to respond

1

8

by pointing out the other pair member (the paired-associate). In case the to-be-paired items were presented consistently in sequence during learning, the associa-tion is one-way: people have more difficulties remembering the association whenthey are cued with the second item (Hulse and Deese, 1967). In contrast, when theitems are presented multiple times in a shuffled sequential order, simultaneously on-screen, or linked by an overlapping context, the formed association is two-way (Hulseand Deese, 1967), indicating that the complete item pair is encoded as one conjunc-tive representation. The paired-associate learning task is the paradigm of choice tostudy human associative memory, due to its relatively simplicity and resemblance tothe type of learning humans are engaged in during everyday life. Items are tradition-ally words (Hulse and Deese, 1967), but can take the form of many types of stimuli,spanning multiple sensory domains. In the experimental chapters of this disserta-tion, I employ different versions of the paradigm. I exclusively used visual stimuli,but differing in complexity across the experiments. In Chapter 2 we used categor-ical pictures of houses, faces, and human bodies, with the benefit that the neuralprocessing pathways of these stimuli are well-understood (Felleman and Van Essen,1991). In Chapter 3, we designed colorful fractal-like circles, with minimal prior asso-ciations, in order to track the emergence of associative memories. In Chapter 4 weemployed a special version of the paired-associate learning task with object stimuli(see Box 5). Objects allow for rich encoding of associations and facilitate encoding bypromoting people to create stories and visualizations (Standing et al., 1970). In sum,the paired-associate learning paradigm is an essential tool for analyzing associativememory and therefore plays a key role in the experiments presented in this thesis.

1.3.Memory representationsThe brain stores information as lasting memory traces. Although the idea that mem-ories are realized through physical alterations of the brain dates back to Plato (Plato,Theaetetus), an exceptionally adequate framework is provided by Richard Semon⁴.Semon viewed a memory trace as ”the enduring though primarily latent modifica-tion in the irritable substance produced by a stimulus” (Semon, 1921). According toSemon, a memory trace is formed when ”all simultaneous excitations form a con-nected simultaneous complex of excitations which, as such acts engraphically, that isto say leaves behind it a connected and to that extent, separate unified engram com-plex” (Semon, 1923). In modern wordings: a memory trace can be considered animprint, represented by an ensemble of myriad neurons (Tonegawa et al., 2015). Buthow does this engram complex of neurons represent information? An intuitive ac-count of the neurobiological basis for a memory trace is given by philosopher RenéDescartes. Descartes figured an analogy with a piece of tensioned linen cloth, which

⁴Sometimes referred to as the ”neglected pioneer of memory science” (Schacter, 2001)

1

9

is briefly punctured with an array of sharp needles. The tension on the cloth causesmost holes to close. However, during a subsequent attempt, the needles would passthrough the cloth more swiftly, since the holes will reopen easily (Figure 1.2).

Figure 1.2: Descartes’ analogy for associativememory.Needle punctures in a linen cloth represent the physical trace of a memory after encoding. The holesare reopened more easily after the first engraving, just as an episodic memory can be triggered andrelived. Figure from (Descartes, 1649).

Similarly, in the brain ”spirits coming upon the pores enter therein more readilythan into others” (Descartes, 1649). In modern wordings: nerve impulses (i.e. Carte-sian spirits) are transmitted more easily by those synapses (i.e. Cartesian pores)that have relayed similar signals at an earlier point in time. This notion capturesthe idea of synaptic plasticity, postulated by Canadian neuropsychologist DonaldHebb: neurons that fire together, wire together⁵ (Hebb, 1949). These changes in re-sistance across synapses, that connects neurons of the engram complex, provide thebiological basis of the memory trace. The idea that memory is imprinted resonateswith the view that memory comprises a chain or network of associations betweenelements (Boring, 1950). Associated elements are engraved together as they are ex-perienced together. In addition, previously encoded engram complexes that are as-sociated with the experience, are reactivated, and they in turn activate yet anothercollection of engrams. The resulting memory network is thus continually shapedand expanded by learning (Milivojevic and Doeller, 2013). A single engram complexis vastly distributed across the brain and in principle never exclusive to a single re-gion in the brain (Lashley, 1950). For instance, in the case of an example episodicmemory, each individual sensory feature is predominantly represented in a differ-ent brain region: faces are represented in the visual system, voices in the auditorysystem and the smell and taste of the food in the gustatory and olfactory systems

⁵This well-known summary phrase was coined by Siegrid Löwel, based on Hebb’s work, but ignores theimportance of temporal precedence in cell firing: neuron A needs to fire just before neuron B in orderto capture their causal relation.

1

10

(Rissman and Wagner, 2012). But where in the brain do these different sources of in-formation come together? Where are separate elements represented in conjunctionand bound into a coherent memory?

1.4. The hippocampusIn 1957, a patient known as H.M experienced a severe case of anterograde amnesiaafter a surgical procedure aimed to treat his intractable epilepsy. Although someof his past memories remained intact after the surgery, he was completely unableto encode new memories (Scoville and Milner, 1957). During the surgery, doctorshad removed a structure called the hippocampus, bilaterally, from H.M.’s brain, thesuspected locus of his epileptic seizures. The case of H.M. serves as the textbook ex-ample highlighting the critical role of the hippocampus for episodic memory. Sincepatient H.M., evidence from additional hippocampal lesion studies, both in animals(Bunsey and Elchenbaum, 1996) and humans (Stark, 2002) has corroborated the ob-served crucial role of the brain region for memory. In addition, modern non-invasiveneuroimaging techniques, such as functional magnetic resonance imaging (fMRI)(see Box 2), have provided insights on memory functions of the hippocampus in thehealthy human brain (Eldridge et al., 2000; Davachi et al., 2003; Rissman andWagner,2012).

Box 2: Functional magnetic resonance imaging

Functional magnetic resonance imaging (fMRI) is the key neuroimaging techniquethat sparked the brain mapping era of the past two decades (Huettel et al., 2008). Bycombining strong magnetic fields and radio wave technologies, fMRI enables us todetect changes inmagnetic properties of blood. As neurons need energy and oxygento function, activity in a region of the brain causes an increase in metabolism, an in-crease in oxygen uptake and subsequently a dynamic regulation of blood flow. Thishemodynamic response is relatively slow compared to the actual timescale of brainactivity: resupply of oxygenated blood peaks around six seconds after the neuronalactivity burst has occurred. As oxygenated blood has different magnetic propertiesthan oxygen-depleted hemoglobin, we are able to measure a Blood OxygenationLevel Dependent (BOLD) effect, allowing us to indirectly infer brain activity (Huet-tel et al., 2008). BOLD effects are reflected in the pixel intensity of an fMRI imageand can be collected over time, yielding a time series with a temporal resolution ofseveral seconds. The technique offers a reasonable spatial resolution: we can accu-rately localize brain activity on a millimeter scale and map the activity of an array of3D pixels, called voxels, across the entire brain. However, caution is needed wheninterpreting fMRI activity maps: the inferential distance between actual changes inbrain activity and the cognitive function under study is relatively lengthy, as themea-

1

11

surements indirectly reflect activity in the brain: neuronal firing, glucose and oxygenconsumption, metabolism increase, increased blood supply, oxygen dispatch, differ-ential magnetic signature of hemoglobin, measurement and statistical modeling con-stitutes the inferential chain from images to brain activity (Huettel et al., 2008). Asa consequence, the inferences drawn from BOLD effects are uncertain. Simultane-ous fMRI and electrical recordings in animals have shown that BOLD correlates withelectrical signals linked to the input and intracortical processing in a brain region,but not the spiking output of neurons (Logothetis et al., 2001). Therefore, BOLD ef-fects might also reflect inhibition of a cognitive function. In addition, conventionalfMRI research aims to isolate brain activity associated with a cognitive function bycontrasting experimental conditions: in one condition, the person being scanned isengaged in the cognitive activity of interest, and in the other condition not (Huet-tel et al., 2008). The ubiquitous nature of memory renders this traditional approachproblematic, since we might not be able to isolate the cognitive activity of inter-est and ensure that there is no interaction with other psychological processes. Forthese reasons, clever experimental designs and alternative ways of analyzing fMRIdata are required to further our understanding of associative memory. Because ofits relatively high spatial resolution and low invasiveness, I use fMRI to investigateassociative memory in experimental Chapter 2 and Chapter 3, in combination withstate-of-the-art analysis techniques (see Box 3). During these experiments, partici-pants were positioned inside the fMRI scanner in a supine position and performedan experimental task, displayed on a back-projected screen through a mirror. Theyresponded according to the task instructions by pressing buttons. Although theseconditions limit the ecological validity of fMRI, themethod provides us with a uniqueview inside the functioning healthy human brain.

The hippocampus is an elongated brain structure, located deep inside the tem-poral lobe underneath the neocortical surface (Andersen et al., 2006). When ex-sected, the structure bears resemblance to a seahorse, and is therefore named afterthe horse-sea-monster in Greek (Turgut and Turgut, 2011). The hippocampus is anarchicortical structure, a phylogenetically old brain region present in all vertebrates,including humans. The region consists out of two interlocked components: the den-tate gyrus (DG), a structure containing granule cells with high rates of neurogenesis,and the cornu ammonis (CA) region, named after the ram’s horns of the Egyptian de-ity Amun (Figure 1.3). The CA comprises multiple subfields, which are numbered oneto four, each with their own specialized cell types and function. In its entirety, thehippocampus basically forms an input-output loop (Buzsáki, 1996). Several anatom-ical pathways connect the subregions to each other and to the rest of the brain viathe entorhinal cortex (EC), a neighbouring cortical structure acting as a relay station(van Strien et al., 2009). The cardinal input to the hippocampus is provided throughthe perforant pathway, consisting of direct projections from the EC to granule cells

1

12

of DG. From DG, mossy fibers project to pyramidal cells of subfield CA3. Via so-called Schaffer collaterals passing through the CA1 subfield, the CA3 projects to theoutput station of the hippocampus, the subiculum, and back to the EC. This unidi-rectional pathway is called the trisynaptic loop. Apart from this circuit, several otherpathways exist that connect subregions of the hippocampus (Figure 1.3).

hippocampus

entorhinal cortexentorhinal cortex

hippocampus

A B

Figure 1.3: Hippocampus anatomy.A) Location of the hippocampus and neighbouring entorhinal cortex in the medial temporal lobe. B)Cross-section of the hippocampus, revealing the resemblance of the CA subregions to a ram’s horns.Input to the hippocampus enters via the entorhinal cortex to the dentate gyrus, looping back to theentorhinal cortex via the CA subregions. Figure adapted from (Moser and Moser, 2016).

Together, these connections stretch the entire longitudinal axis forming manyparallel processing circuits of fiber grids and functional gradients (Navarro-Schröder,2016). The specialized anatomy of the hippocampus provides some clues about itsfunction: encoding new information and retrieving stored memories. Incoming per-ceptual information enters the hippocampus via the DG. The DG has been put for-ward as region capable of distinguishing very similar input patterns. For example,two representations of slightly different but similar objects, for instance two apples,are orthogonalized by DG to be able to separate them in memory. These orthogo-nalized representations are forwarded to CA3 (Treves and Rolls, 1994). The recurrentconnections of CA3 allow the subregion to act as an autoassociator, completing therepresentation of partial input patterns (McNaughton and Morris, 1987). Throughthis autoassociative mechanism, hippocampal representations are subject to attrac-tor dynamics: inputs are transformed by a logistic function and thus biased to astable state (Wills, 2005; Steemers et al., 2016). For example, when you perceive thesecond apple, the representation of the first apple is recalled from memory, since itis the most similar attractor state. The CA1 region has been put forward as a com-

1

13

parator, that attempts to reconcile perceptual input from the EC with mnemonicoutput from CA3, and feeds back potential mismatch signals (Vinogradova, 2001;Chen et al., 2011). If we apply this idea to our earlier example of the two apples, CA1would detect the slight differences between the apples and forward the mismatchinformation to other brain regions.

But how does the hippocampus store memories? In 1973, the phenomenon oflong-term potentiation was discovered in DG (Bliss and Lømo, 1973). This findingprovided the first evidence for synaptic plasticity, the cellular basis of memory. Twoyears earlier, Nobel prize winner John O’Keefe made the seminal discovery of spa-tially selective neurons, called place cells (O’Keefe and Dostrovsky, 1971), using elec-trophysiological recordings in freely-moving rodents. These cells only fire when ananimal is at a certain location, and are responsive to spatial cues in the environ-ment. Each cell fires at a different location, together effectively forming a map ofthe animal’s environment. Hippocampal place cells enable an organism to encodeevents taking place at certain locations in the environment and form a cognitivemap that integrates the what and where (Tolman, 1948; O’Keefe and Nadel, 1978).In addition, electrophysiological recordings have revealed other types of hippocam-pal cells that code for non-spatial aspects of an episode, such as time (Manns et al.,2007; Pastalkova et al., 2008), remembered objects (Manns and Eichenbaum, 2009)and even conceptual information (Quiroga et al., 2005). What these cell types havein common is that they represent conjunctions of features that characterize an envi-ronment, episode or concept. Apart from the mentioned electrophysiological work,recent advancements in neuroimaging analysis techniques have provided insightsin these conjunctive representations in the healthy human brain (Chadwick et al.,2010; Shohamy and Wagner, 2008; Staresina et al., 2013; Davachi et al., 2003; Kuhlet al., 2013; LaRocque, 2013; Azab et al., 2014; Copara, 2014; Rissman and Wagner,2012; Milivojevic et al., 2015; Collin et al., 2015) (see Box 3). Similarly, in experimentalChapter 2 of this thesis, my coauthors and I looked into the nature and distribu-tion of conjunctive representations across the brain, with a special focus on the hip-pocampus. However, it remains unclear how these conjunctive representations areformed and whether we can decode newly formed associations from non-invasivelymeasured brain activity patterns. To answer this question, in Chapter 3, we trackedthe emergence of conjunctive representations in the hippocampus as a function oflearning.

1

14

Box 3: Investigating brain representations

In response to a stimulus, some neurons in the brain are very active, some showsporadic activity and others remain silent. Across such an assembly of cells, a spa-tiotemporal pattern of activity emerges, that represents some piece of informationrelated to the stimulus (Hebb, 1949). With fMRI however, we are unable to mea-sure individual cells and rather pick-up the activity from cubic-millimeter-sized vox-els. These signals provide us with an indirect measure of the activity of a popula-tion of neurons (see Box 2). Fortunately, small biases in activity across voxels resultin a multivariate spatial information pattern, allowing us to investigate brain repre-sentations (Kamitani and Tong, 2005). In recent years, multivariate pattern analysishas become increasingly popular in the neuroimaging community (Haynes and Rees,2006). Rather than looking at the average level of activity in a brain region, we an-alyze the activation pattern across voxels, to model how the brain encodes certaintypes of information. In a next step, these encoding models can be used in reversefor the purpose of brain-reading: decoding information from patterns of brain activ-ity. These models can be the result of sophisticated machine learning techniques(Varoquaux and Thirion, 2014), or simply looking at the similarity structure of pat-terns across different stimuli or experimental conditions. The latter is a techniquecalled representational similarity analysis (RSA) (Kriegeskorte et al., 2006), a methodwhich we adopted in Chapter 2 and Chapter 3. Here, we correlate patterns of activ-ity, to obtain proxy measures of neural similarity between the different experimentalconditions (Figure 1.4).Subsequently, we extract summary statistics from these similarity data to index thetype of represented information. This procedure can be performed on voxel pat-terns from an a-priori region-of-interest or executed iteratively using arbitrarily-sizedspheres across the entire brain. With this so-called searchlight approach, we are ableto assess where in the brain certain types of information are more prominently rep-resented (Kriegeskorte et al., 2006). One drawback of multivariate pattern analysisapproaches is their susceptibility to temporal autocorrelation: patterns close in timeare inevitably more correlated (Haynes, 2015; Hsieh et al., 2014). This property isparticularly problematic if we wish to study changes in representations of time asa function of learning, when scanner noise and motion artifacts are the main cul-prits of temporal autocorrelation. In addition, non-specific brain activations duringthe learning task might obscure subtle changes in brain representations. One wayto overcome these issues is by performing RSA on independent data acquired be-fore and after the learning task. This approach allows us to investigate representa-tional reconfigurations, by comparing the post-learning with the pre-learning similar-ity structure (henceforth referred to as differential RSA) (Visser et al., 2011; Schapiroet al., 2012; Schlichting et al., 2015; Milivojevic et al., 2015; Collin et al., 2015). In Chap-ter 3, we used a combination of this experimental design and multivariate pattern

1

15

analysis to investigate the formation of associative memory representations.

conditions

searchlight

voxel patterns similarity

graph

brain map

Figure 1.4: Representational Similarity Analysis.Neural activity profiles for different experimental conditions (e.g. pictures of faces, houses)are recorded with neuroimaging. For a given region-of-interest, the multi-voxel pattern ofactivity is extracted and cross-correlated. A high correlation between a pair of conditions im-plies high neural similarity (indicated by the dark shades of gray in the similarity matrix). Thesimilarity structure can be visualized as a graph, yielding a structure akin to the theoretical as-sociative memory networks sketched in Figure 1.1. A summary metric may be computed fromthe similarity data, to index the information content of a region. This metric may be assignedto the respective region-of-interest or the searchlight sphere center, in order to obtain a brainrepresentation map.

The discovery of hippocampal conjunctive representations, which bind featuresor elements of an episode memory in particular, have culminated in hippocampalindex theory. In this framework, the hippocampus has been posited to store rep-resentations of episodic memories by encoding pointers to the neocortical repre-sentations of episode features (Marr, 1971; Teyler and DiScenna, 1986; Murre, 1996),analogous to the index of a book or digital search engine. In terms of anatomical con-nectivity, the hippocampus is ideally equipped to perform this function. As outlinedabove, the structure is remarkably well-connected to the rest of the brain via theEC. In addition, connections between the hippocampus and neocortical regions fol-low the general motif of convergence (Papez, 1944): starting from primary sensorycortices, connection pathways converge onto higher-order brain regions, with thehippocampus at the apex (Damasio, 1989; Hoesen et al., 1972). Recent advances inneuroimaging analysis techniques have enabled us to investigate the cross-regionalfunctional interactions in the healthy human brain, while it performs certain cogni-

1

16

tive tasks (Friston, 2011) (Box 4). Using these methods, studies have revealed thatthe hippocampus connects to sensory brain regions during memory encoding andretrieval processes (Staresina et al., 2013; Schlegel, 2013; Schedlbauer et al., 2014).However, it remains unclear whether this convergent functional connectivity profileof the hippocampus is implicated in the retrieval of conjunctive memory representa-tions. In experimental Chapter 2 of this thesis, we employed functional connectivitymethods to assess whether the hippocampus acts as a network hub during memoryretrieval.

Taken together, the hippocampus is a brain structure responsible for binding var-ious pieces of information into a conjunctive representation, which thereby form acoherent memory. The unique neuroanatomy and circuitry of the hippocampus al-lows thismemory to be encoded and subsequently retrieved. These propertiesmakethe hippocampus the prime region-of-interest in our study of the associations thatcharacterize a single episodic memory. But what happens when multiple episodicmemories are related to each other?

Box 4: Investigating brain connectivity

Anatomical connections between brain regions can be investigated in several ways,for instance by injecting neuronal traces or with a technique called Diffusion TensorImaging (Filler, 2009). Besides anatomical pathways, we may also study functionalinteractions between brain regions during rest (Biswal et al., 1997; Raichle et al., 2001;Vincent et al., 2006) that support a given cognitive task. Here, we quantify the statis-tical dependencies between brain regions and look for correlated patterns of activityacross time (Biswal et al., 1997). Here, the assumption is that synchronized brain re-gions are more likely to interact with each other than regions whose activation timeseries are uncorrelated. When performed on whole-brain data, the procedure yieldsan estimate of the amount of functional connectivity between each pair of brain re-gions within the entire network (Figure 1.5). Using concepts from graph theory, wecan now employ network analyses to identify well-connected hub regions, whichare crucial for effective communication (Bullmore and Sporns, 2009). The brain isa complex network, consisting of many subnetworks, called modules (Bullmore andSporns, 2009). So-called hub regions play a core role in connecting these modules.Provincial hubs are central in a local sense and connect regions within the samemod-ule. In contrast, connector hubs link-up the different modules in the brain (van denHeuvel and Sporns, 2013). In experimental Chapter 2, we computed a specializedhub measure that quantifies the diversity of a region’s connections between differ-ent modules (Power et al., 2013). The combination of fMRI connectivity methods andnetwork analyses allowed us to investigate convergence of functional connections

1

17

during associative memory retrieval.

connectorhub

provincial hub

network graph

activity profiles

brain regions

connection matrix

Figure 1.5: Functional brain connectivity.Connectivity between brain regions can be derived from the cross-correlation structure oftheir time series or any other series of repeated measures of activity. The statistical depen-dencies between regions is represented by a symmetric connectionmatrix (middle), indicatingwhich brain regions communicate with each other and the strength of their interaction (hotcolors indicate a strong interaction, while cooler colors indicate a weak interaction). The con-nectionmatrix may be thresholded and subsequently visualized as a network graph with brainregions as nodes and their interactions as edges. The typical brain network graph comprisesmultiple separate functional modules (gray circles) and different types of hub nodes (maroonand ochre).

1.5.Memory integrationThe formation of associations between related events is a vital step in the trans-formation of context-specific episodic memories to generalized abstract knowledge(Eichenbaum et al., 1999; Kumaran et al., 2009; Zeithamova et al., 2012b; Shohamyand Turk-Browne, 2013) (Figure 1.1). By integrating memories that share certain fea-tures or contextual details, we are able to infer relationships between elementsthat were never associated explicitly. Ultimately abstract rules can be learned fromthe resulting regularities across events (Doeller et al., 2005). Taken one step fur-ther, we may hypothesize that memory integration through associative inferenceis the mechanism underlying knowledge acquisition. Unsurprisingly, the hippocam-pus also plays an important role in memory integration (Bunsey and Elchenbaum,1996; Dusek and Eichenbaum, 1997; Zeithamova et al., 2012a; Schlichting et al., 2015;Nagode and Pardo, 2002; Greene et al., 2006; Preston et al., 2004; Collin et al., 2015;Milivojevic et al., 2015), since the mechanism to retrieve memories and encode theirinterrelations is required. But how are actually memories combined to form new

1

18

representations? Computational models and experimental work on associative in-ference (see Box 5) suggests that integration of two separate episodic memoriesis realized through retrieval-mediated learning (Kumaran and McClelland, 2012; Zei-thamova et al., 2012a; Shohamy and Wagner, 2008).

Box 5: The associative inference paradigm

Memory integration can be studied as a variation of paired-associative learning. Inthe resulting associative inference paradigm (Heckers et al., 2004; Preston et al.,2004; Zeithamova et al., 2012b), a person is initially tasked with learning two so-called premise pairs. In this first stage, an item A is paired to item B. Secondly, itemB is paired to item C, thereby creating two premise associations, tagged AB and BC,that both share a common element, B. Through logical deduction, the person is thenprompted to link item A to item C, creating an third association, AC. Crucially, theAC combination has never been observed by the person directly, but is rather in-ferred from the premise associations. In experimental Chapter 4, we employ theassociative inference task to study the neural mechanisms of memory integration.We used clearly recognizable object stimuli, to facilitate one-shot encoding of theassociations through explicit learning. We combined the task with a subsequentmemory analysis, also known as the difference due to memory (Dm) analysis (Pallerand Wagner, 2002). Here, we exploit the results from the memory test to isolateand investigate the encoding-related brain signals associated with memory success.The combination associative inference task with a subsequent memory analysis al-lowed us to assess which types of neural signals during encoding are predictive ofsuccessful memory integration.

To illustrate this principle, let us consider memory of two events with overlap-ping contexts: one past event and one new, being encoded. At the moment whenthe newmemory is encoded, the old memory is pattern-completed in the hippocam-pus, since the overlapping context reactivates part of the old memory engram com-plex (Horner and Burgess, 2014). This new, combinedmemory trace is then encoded,leaving the two memories associated. Next, we might wonder how this link is en-coded. One option is that a completely new unified memory is stored, where bothepisodes are represented together by one engram complex (Kumaran and McClel-land, 2012). An alternative theory posits that each episode is stored separately, butthe shared feature, for instance the context, connects both memories. Studies havefound experimental evidence for both theories (Kumaran andMcClelland, 2012) andefforts have been made to reconcile them in a common framework (Zeithamovaet al., 2012b). In line with this idea, recent neuroimagingwork has demonstrated thatmemories with different event resolutions can coexist in the hippocampus (Collin

1

19

et al., 2015). These findings suggest that the brain stores representations of indi-vidual episodes, as well as unified multi-episode representations which integrateinformation. Ultimately, these associated memories form an interlinked hierarchi-cal memory network, representing information at multiple scales. But how does thehippocampus know which memories need to be integrated?

Experimental evidence suggests that the medial prefrontal cortex (mPFC) is akey brain region involved in memory integration (Preston and Eichenbaum, 2013;Zeithamova et al., 2012a; Schlichting et al., 2015). Research in animals (DeVito et al.,2010) and patient studies (Koscik and Tranel, 2012) have shown that lesions of themPFC lead to significant decreases in performance on various memory-integrationtasks (see Box 5). The mPFC envelops the entire medial wall of the prefrontal cortexand is particularly well-developed in primate species, including humans (Öngür et al.,2003; Wise, 2008). The ventral part is of particular interest when studying memoryand decision-making, as it is well-connected with the nearby limbic system, includ-ing the hippocampus. Hitherto, there is no consensus regarding the demarcationof the ventral mPFC, but in general, the subregion is thought to encompass Brod-mann areas (BA) 10, 14, 25 and 32, and include parts of BAs 11, 12, 13 and 24. Unlikemost neocortical brain regions, the mPFC receives direct monosynaptic inputs fromthe hippocampus (Jay and Witter, 1991). Back-projections from the mPFC to the ECand an additional subcortical route via the thalamus to the hippocampus (Figure 1.6)enable the mPFC to reciprocally interact with the hippocampus.

medial prefrontal cortex

A B

10

25

32

10

11 14

32

9

14

10

8

10 24

thalamus

Figure 1.6: Medial prefrontal cortex anatomy.A) The ventral part of the prefrontal medial wall receives direct input from the hippocampus. Back-projections via the entorhinal cortex and thalamus enable a bidirectional flow of information betweenthe mPFC and the hippocampus. B) The mPFC comprises several cytoarchitectonically distinct Brod-mann areas. White borders indicate a proposed further subdivision of these BA regions, based on theirfunction (Wise, 2008). Figure adapted from (Wallis, 2011).

1

20

In accord with the key role of the mPFC in memory integration and its strongconnectivity with the hippocampus, fMRI studies have found increased functionalconnectivity between the hippocampus and mPFC during memory integration (Ku-maran et al., 2009; Zeithamova et al., 2012a; van Kesteren et al., 2010). Throughfunctional interactions, the mPFC might guide the selection of inputs to the hip-pocampus or modulates retrieval of memories from the hippocampus (Barron et al.,2013; Preston and Eichenbaum, 2013). This notion is in line with experimental evi-dence suggesting that the mPFC accumulates contextual information of memories(Tse et al., 2007; Hyman et al., 2012). Moreover, these context-decoding neuronalensembles in the mPFC have been found to rapidly switch dependent on task con-tingencies (Durstewitz et al., 2010). The dynamic switch between contexts mightbe related to conflicts arising from different sources of perceptual and mnemonicinformation. The mPFC may help to resolve these conflicts, during the encodingand retrieval of memories. These operations are important for memory integrationthrough retrieval-mediated learning (Zeithamova et al., 2012a): the system needs tobe tuned to retrieve the previously stored memories that share context with the in-coming new information. Subsequently, conflicting information from the old mem-ory and new memory need to be tagged and resolved, and a new integrated mem-ory needs to be encoded. Accordingly, the interactions between the mPFC and thehippocampus are vital for memory integration. But what is the neurophysiologicalmechanism by which these regions communicate?

1.6. Theta oscillationsThe prime candidate mechanism for hippocampal-prefrontal interactions are thetaoscillations (Hyman et al., 2005). Theta is a type of rhythmic brain activity thatis prominent in both brain regions. Large amplitude theta oscillations (frequencyaround 5 Hz in humans, 6 to 10 Hz in rodents) are characteristic for the hippocam-pus (Buzsáki, 2002; Jacobs, 2014) (Figure 1.7).

In general, theta oscillations reflect the online state of the hippocampus(Buzsáki, 2002). When theta emerges, the hippocampus is ready to process incom-ing signals and output stored information. Theta is prevalent during the awake ac-tive states of an animal, for instance when it is navigating its environment, attendingstimuli, or making decisions (Grastyan et al., 1959; Klemm, 1972; Vanderwolf, 1969).In addition, theta is observed during REM sleep, an active dream state during whichawake behavior is simulated (Jouvet, 1969). What all these behaviors have in com-mon, is that information needs to be encoded and retrieved. This regularity indicatesthat theta fulfills a key role for memory (Buzsaki and Moser, 2013). Coherent thetaoscillations can be observed across all subregions of the hippocampus. The medialseptal region plays an important role in the generation of theta. Neurons locatedin the septum have been found to rhythmically inhibit pyramidal cells in the hip-

1

21

250 μV

100 ms

A

B

Figure 1.7: Hippocampal theta oscillations.A) Local field potential, recorded from the rat hippocampus with an intracranial electrode. B) Filteredderivative (14 Hz low-pass) of A. Note the clearly visible brain rhythm, with a period of around 160milliseconds. Figure adapted from (Dragoi et al., 1999).

pocampus and ablation of medial septum abolishes theta oscillations (Stewart andFox, 1990). Moreover, theta band activity is crucial for synaptic plasticity (Hymanet al., 2003), the cellular process underlying memory formation. The rhythmic inhi-bition and excitation of neuronal activity facilitates long-term potentiation, causinga change in synaptic weights that allow the encoding of information. Furthermore,theta provides a mechanism for spatiotemporal coding of information (O’Keefe andRecce, 1993; Buzsaki, 2005). After each recurring inhibitory theta phase, the most ex-citable neurons in an engram complex discharge first, allowing for a cellular narrativeof information in time and space. Lastly, synchronized theta oscillations across re-gions provide a time window for effective communication (Fell and Axmacher, 2011;Fries, 2005). Throughout the entire neocortex, neuronal activity is clocked to thetaoscillations (Siapas et al., 2005; Canolty et al., 2006; Sirota et al., 2008), making the5 Hz brain rhythm an ideal candidate mechanisms for facilitating long range com-munication between brain regions. Accordingly, studies using human intracranialelectrophysiological recordings (Lega et al., 2012; Watrous et al., 2013; Rutishauseret al., 2010) and non-invasive methods, such as EEG and MEG (see Box 6), have im-plicated theta oscillations in amultitude ofmemory processes (Guderian et al., 2009;Cornwell et al., 2008; Guitart-Masip et al., 2013; Kaplan et al., 2014; Riggs et al., 2009;Fuentemilla et al., 2014; Staudigl and Hanslmayr, 2013).

Box 6: Magnetoencephalography

Magnetoencephalography (MEG) is used to measure the electrical activity of thehealthy human brain (Baillet et al., 2001). This non-invasive technique is similar toelectroencephalography (EEG), but sensitive to the minute magnetic fields, gener-

1

22

ated by synchronized neuronal activity, instead of the associated electrical currents.A helmet of superconducting sensors, called magnetometers, is used to pick-upmag-netic fields originating from neuronal population (Figure 1.8).

sensor array pyramidal neurons electrical current magnetic field

Figure 1.8: Basis of themagnetoencephalography signal.Left: The sensor array of magnetometers is aligned to the cortical surface of the brain. Middle:pyramidal neurons generate electrical currents when active (right). A significantly large groupof synchronously active neurons can give rise to measurable magnetic fields. Pyramidal neu-rons that are situated perpendicular to the cortical surface, for instance in the cortical sulci,generate magnetic fields that are best captured by the external sensors. Figure adapted fromwww.humanconnectome.org

MEG is used to measure signals which are directly related to electrical activity, as op-posed to fMRI BOLD effects. Thus, MEG provides measures of brain activity with ashorter inferential distance than fMRI (see Box 2) andmoreover with a very high, mil-lisecond time resolution. The fast temporal dynamics of the brain that MEG revealscan be used to to investigate the role of brain rhythms, such as theta oscillations.In addition, MEG can provide more detailed connectivity measures than fMRI (seeBox 3), by looking at the amount of amplitude and phase synchronization of oscil-latory signals in different brain regions, with or without a lag in time (Siegel et al.,2012). The major advantage of MEG over the more widely-used EEG, is its spatialresolution: since magnetic fields are not distorted by tissue conduction effects, thespatial resolution of MEG ranges from several centimeters to several millimeters, de-pending on depth of a source and movement artifacts. Although the hippocampusis a deep structure, a combination of very specific a-priori hypotheses and advancedsource reconstruction techniques may be exploited to pick-up these deep-sourcesignals (Dalal et al., 2013; Stephen et al., 2005; Attal and Schwartz, 2013). In experi-mental Chapter 4, we employ such an approach to investigate interregional couplingof theta oscillations originating from deep sources.

Taken together, theta oscillations may provide the mechanism by which the hip-pocampus and mPFC are able to integrate memories. However, this hypothesis re-

1

23

mains untested. In Chapter 4, we used MEG to investigate the role of theta os-cillations during memory integration, and specifically their key role in facilitatingcommunication between the hippocampus and prefrontal cortex.

1.7. Thesis outlineIn this dissertation, I present my investigations on the neural mechanisms of asso-ciative learning. I will answer the following overarching question: how are we ableto encode, store and retrieve associative information, which enables us to remem-ber episodes from our lives and ultimately acquire knowledge about our world? Inthe three experimental chapters of this thesis, I have approached this question fromtwo different angles. Firstly, I looked into the nature of associative representations:how the brain connects memories. Secondly, I investigated brain network dynamics:how memories connect the brain.

Specifically, in Chapter 2, we investigated the convergence of mnemonic infor-mation in the hippocampus. Since the early 1970s, influential computational modelshave proposed the existence of so-called convergence zones in the brain (Marr, 1971;Damasio, 1989; McClelland, 1994). These brain regions are thought to house associa-tive representations, which bind distributed cortical elements of an episode such asplaces, objects, events, and people into rich coherent memories. Therefore, we hy-pothesized that a convergence zone can be identified by simultaneously testing (i) itsability to represent conjunctive mnemonic information and (ii) its widespread hub-like interaction with other brain regions. In particular, the hippocampus has beenput forward as the prime candidate acting as a convergence zone for episodic mem-ory. Surprisingly, empirical evidence for the specialized role of the human hippocam-pus in mnemonic convergence is scarce, since previous work has focused on repre-sentational and network properties in isolation: the vast majority of neuroimagingstudies on memory pursued a region-of-interest approach, thus neglecting the net-work perspective, whereasmost connectivity studies ignored task-relevant represen-tations. Therefore, we aimed to answer the following question: is the hippocampusa mnemonic convergence zone based on its connectivity and representational prop-erties? To answer this question and test our overarching hypothesis, we leverageda novel combination of two analysis techniques, tailored to gauge the two key prop-erties of a convergence zone. We applied a whole-brain representational similarityanalysis in combination with a graph-theoretical network approach to functionalmagnetic resonance imaging (fMRI) data acquired during an associative memorytask. Next, we inspected the overlap of conjunctive memory coding (using across-voxel fMRI pattern correlation as a proxy for neural similarity) and hub-like networkattributes in the hippocampus. The findings presented in this chapter provide evi-dence for mnemonic convergence in the hippocampus. By simultaneously assessingrepresentational and network metrics, we shed new light on the neural mechanisms

1

24

in the hippocampus, in line with its posited crucial role in binding distributed infor-mation into integrated mnemonic representations.

In Chapter 3, we further studied conjunctive representations and how they areformed in particular. Recent functional magnetic resonance imaging studies on im-plicit temporal regularity and inference learning have shown group-level representa-tional reconfigurations, by comparing neural activity patterns before and after learn-ing (Visser et al., 2011; Schapiro et al., 2012; Schlichting et al., 2015; Milivojevic et al.,2015; Collin et al., 2015). However, interrogating these memories and assessing theirrepresentational geometry still remains challenging. In this chapter, we aimed to an-swer the following question: can we track the emergence of associative representa-tions, both across a group of participants and on the level of an individual? To answerthis question, we employed differential RSA (see Box 3) to investigate the neural simi-larity structure of simple visual stimuli before and after participants learned randomstimulus pairs. We hypothesized that after learning, representational similarity ofthe associated items would increase relative to the non-associated material. Subse-quently, we assessed the predictive value of the neural similarity structure to identifyassociations and read-out memories from an individual’s neural data. The findingsreported in this chapter present a litmus test for the differential RSA approach, andprovide methods to evaluate reconstructed individual memory networks.

Next, we addressed the neuralmechanisms of cross-episode integration ofmem-ories in Chapter 4. Recently, models of hippocampal-prefrontal interplay subservingmemory integration have been proposed (Preston and Eichenbaum, 2013), but theunderlying neurophysiological mechanisms are poorly understood. Previous elec-trophysiological studies studies on other memory functions suggest theta band os-cillatory activity as a likely candidatemechanism (Benchenane et al., 2010; Andersonet al., 2010), but no study to date has shown that theta is involved in memory inte-gration and inferential reasoning in humans. Therefore, in this chapter, we aimed toanswer the following question: what is the role of theta oscillations in memory inte-gration? We leveraged the superior temporal resolution of magnetoencephalogra-phy (MEG) to answer this question and recordwhole-brain oscillatory activity duringa memory-integration task. We used novel, advanced source reconstruction meth-ods to estimate hippocampal oscillatory signals, and showed that both amplitudeand coupling strength of theta oscillations in hippocampal and medial prefrontalsources predicts successful memory integration. By directly relating memory-baseddecision-making source-level theta oscillations, the findings presented in this chap-ter demonstrate a neurophysiological mechanism by which the medial prefrontalcortex and hippocampus support the integration of disparate memories.

In Chapter 5, I summarize and combine the results of the experimental chapters,and describe the advancements made in this thesis on several aspects, includingidentifying anchors for future research and development. Firstly, I highlight how the

1

25

findings presented in this thesis further our neuroscientific understanding of mem-ory. Secondly, I delineate the methodological advancements. In the last section ofthis thesis, I speculate on the potential applications for society of the key observa-tions andmethods used in this thesis in two domains: clinical and education. Finally,I present the concluding remarks.

2Mnemonic convergence in the human

hippocampus

Alexander R. Backus, Sander E. Bosch, Matthias Ekman, AlejandroVicente Grabovetsky, Christian F. Doeller

This chapter has been published as: Backus, A.R.*, Bosch, S.E.*, Ekman, M., Vicente Grabovetsky, A.,Doeller, C.F. (2016). Mnemonic convergence in the human hippocampus, Nature Communications,7(11991):1–9, doi: 10.1038/ncomms11991

A different subset of the data from this project resulted in a twin chapter, which has been published inthe dissertation of Sander E. Bosch as: Bosch, S.E.*, Backus, A.R.* Doeller, C.F. (2016) Memory represen-tations shift from hippocampus to medial frontal cortex through memory consolidation, Reactivatingmemories in hippocampus and neocortex, Chapter 5, 83–98, ISBN 978-94-6284-042-3

* denotes equal contributions.

27

2

28

2.1. IntroductionEpisodic memories entail a rich set of different features, such as the place where anevent occurred (for example, the local bakery), the people encountered (for exam-ple, a teacher from our children’s school), the content of a conversation (for exam-ple, the upcoming Christmas party at school) and when it took place (for example,Wednesday afternoon). An important aspect of memory formation is the conver-gence of such separate elements onto a conjunctive representation (Tulving andWatkins, 1975; O’Reilly and Rudy, 2001). This convergence of information is crucialnot only for simple associations between stimulus features, but just as much for thebinding of relationships between places, people, objects and events into complexepisodic memories. But how does the brain implement mnemonic convergence?Computational models of memory have hypothesized for a long time that special-ized modules, so-called convergence zones, exist in the brain (Marr, 1971; Damasio,1989; McClelland, 1994; O’Keefe and Nadel, 1978; Eichenbaum, 2000). These zonesare characterized by two key properties: conjunctive coding and a high degree ofinterconnectivity with other brain regions. Although the existence of convergencezones is widely acknowledged, there is as of yet limited evidence for their neuralunderpinnings.

A prime candidate for mnemonic convergence is the hippocampus, a brain re-gion that is thought to index the cortical elements of an episodic memory repre-sentation (Eichenbaum, 2000; Teyler and DiScenna, 1986; Stark and Squire, 2001)by means of conjunctive coding (Marr, 1971). In line with this idea, several theorieshave posited a key role for the hippocampus in binding item and context informa-tion and binding of discontiguous elements (Davachi, 2006; Milivojevic et al., 2015).Experimental evidence from studies using electrophysiological recordings (Moitaet al., 2003; Wood et al., 1999) and functional magnetic resonance imaging (fMRI) inhumans (Chadwick et al., 2010; Shohamy and Wagner, 2008; LaRocque, 2013; Azabet al., 2014; Copara, 2014; Horner et al., 2015) support conjunctive representations inthe hippocampus using a wide array of experimental tasks. However, many of thesestudies have pursued a region-of-interest approach, thus neglecting the networkperspective. In parallel, the convergent connectivity profile of the hippocampus hasbeen traditionally examined using neuronal tracer techniques in animals (Hoesenet al., 1972; Felleman and Van Essen, 1991) and neuroimaging connectivity methodsin humans (Lewis et al., 2009; Vincent et al., 2006; Watrous et al., 2013). Althoughsome studies investigate network properties during cognitive tasks (Schlegel, 2013;Ekman et al., 2012; Schedlbauer et al., 2014; Navarro-Schroder et al., 2015; Ritcheyet al., 2014), many connectivity studies focus on the entire (often rodent or monkey)brain at rest, ignoring the relationship between brain connectivity and task-relevant,regionally specific representations. Thus, surprisingly, the two key properties thatdefine a convergence zone, namely conjunctive representations (hereafter referred

2

29

to as conjunctiveness) and interconnectivity with other brain regions (hereafter re-ferred to as hubness), have hitherto been only studied in isolation in the humanhippocampus.

Here we investigate whether the hippocampus is a convergence zone and testthe prediction that the hippocampus plays a special role in associative binding. Weuse a simple associative learning paradigm and fMRI techniques, in combinationwith two analysis approaches to simultaneously gauge the two key properties of aconvergence zone: we employ representational similarity analysis (RSA) (Kriegesko-rte et al., 2006) to assess neural representation of conjunctiveness in regional mul-tivoxel patterns and adopt a graph-theoretical network approach (Bullmore andSporns, 2009) to quantify hubness from the functional connectivity data duringmemory retrieval. Subsequently, we assess the overlap of these two neural metricsas a marker of mnemonic convergence. Importantly, we employ whole-brain anal-yses to investigate a region-specific question: is the hippocampus a convergencezone, characterized by a combination of both conjunctiveness and hubness?

2.2. ResultsParticipants performed a paired-associate retrieval task in theMRI scanner after hav-ing learned the associations between pairs of grayscale images of faces, houses andfaceless bodies (Figure 2.1). All participants (N = 25) were able to remember the asso-ciations with high accuracy (average performance: 84.6% correct responses, s.e.m.= 2.0%, average d-prime: 1.61, s.e.m.: 0.07).

Conjunctive representations during memory retrievalTo test whether the human hippocampus fulfills the criteria of a convergence zone,we specifically aimed to detect overlap of conjunctiveness and hubness. We opera-tionalized conjunctiveness as the amount of information about specific memory as-sociations in patterns of fMRI activity. To this end, we alternated temporal order ofcue and paired-associate instances already during learning, to have participants cre-ate one conjunctive representation for each pair of stimuli, independent of their or-der. We then systematically assessed the presence of these conjunctive representa-tions in spherical regions surrounding a single voxel (search lights), using RSA. Specif-ically, we applied a representational similarity contrast where we expected higherneural pattern similarity when comparing instances of the same association relativeto comparing different associations (that is, associative similarity, Figure 2.2A). In ad-dition, we imposed a perception penalty on this contrast by excluding perceptuallysimilar comparisons and thereby emphasizing perceptually dissimilar comparisons.As a result, category-related perceptual contributions to pattern similarity were pe-nalized, maximizing the sensitivity of our analysis to detect conjunctive mnemonicrepresentations (see section 2.4 and Figure 2.2A for details). The final contrast re-

2

30

?

0.2s

1-11s

0.2s

0-0.6s

1-11s

encoding

session

cue

probe

recall task

in scanner

?

?

Figure 2.1: Experimental procedure and trial structure.Top: participants learned the associations between grayscale pictures depicting either a face plus body,scene plus face or body plus scene, during an initial encoding session. Subsequently, participantsretrieved these associations in the scanner. Note that in the actual experiment, category icons andpair numbers (used here for illustration purposes) were replaced by photographic stimuli, as describedin the section 2.4. Bottom: each retrieval trial comprised a cue, a retrieval phase of variable length andthe presentation of a match or non-match probe stimulus (bottom). Participants indicated whetherthe probe matched the paired-associate by button press. Trials were separated by a variable inter-trialinterval.

sulted in a conjunctiveness score for each individual voxel (Figure 2.2C left panel).As predicted, the hippocampus showed a significant conjunctiveness effect (peakMontreal Neurological Institute (MNI) coordinates: x,y,z = [ -30, -16, -14], T₂₄ = 3.59, p= 0.045 small-volume family-wise error (FWE) corrected using Threshold-Free Clus-ter Enhancement, see section 2.4 for details). We found no significant differences in

2

31

other brain regions in a follow-up analysis (p > 0.17 whole-brain FWE-corrected, seeFigure 2.3A).

Network centrality during memory retrievalIn parallel, we employed a whole-brain graph theoretical analysis to probe neuralnetwork dynamics during associative memory retrieval. To this end, we computedbeta time-series correlations between all grey matter voxels in the brain (Rissmanet al., 2004) and summarized the connectivity profile of each voxel into a hubnessscore (Power et al., 2013) (Figure 2.2B, see section 2.4 for details). We used the par-ticipation coefficient as our hubness metric, which quantifies the importance of agiven node (that is, voxel) for interactions between subnetworks (Power et al., 2013;Guimera et al., 2005). Nodes participating in multiple subnetworks (so-called con-nector hubs) are likely integrating different types of information across distributedbrain regions and function as convergence zones (Power et al., 2013). In addition,the participation coefficient provides a more sophisticated and robust index of hub-ness than traditional measures, such as degree centrality (Hannula and Ranganath,2008). To obtain a task-related measure of hubness for each voxel, participation co-efficients during memory retrieval were contrasted with rest intervals (Figure 2.2Cmiddle panel), in the absence of head displacement differences between task phases(see Figure 2.4 and section 2.4 for details). In line with our predictions, the hip-pocampus showed a significant retrieval-related hubness effect (peak MNI coordi-nates: x,y,z = [28, -14, -22], T₂₄ = 3.75, p = 0.009, small-volume FWE-corrected, seeTable 2.1 for a list of other conjunctiveness cluster peaks and their hubness scoresfor comparison). We found no significant hubness differences in other brain regionsin a follow-up analysis (p > 0.42 whole-brain FWE-corrected, Figure 2.3B, see Fig-ure 2.5 for the participation coefficient map from the rest interval only). In addition,to corroborate the participation coefficient results, we repeated the analysis with adifferent hubness metric, eigenvector centrality, on alternatively preprocessed data(see section 2.4 for details). We observed a similar hippocampal effect (peak MNIcoordinates: x,y,z = [32, -18, -16], T₂₄ = 3.56, p = 0.046, small-volume FWE-corrected,Figure 2.6).

2

32

p = 0.0004p = 0.0135

0 0.510-4

100

Dice coe!cient

pro

b

0 0.0610-4

100

relative overlap size

pro

b

ne

ura

l sim

ilari

ty (

a.u

.)

comparisons

Me

an

pa

rtic

ipa

tio

n

coe

!ci

en

t

ITI recall0.4

0.6*

**

***

**

-16

L

-30

-14

28

-14

-30-30

T

40

voxels

voxe

ls

betas

voxe

ls

same di"erent

association

cue and

associate

category

same

di"erent

dissimilar

similar

not compared

not compared

cue associate

versus

A

B

C

D

E

2

33

Figure 2.2: Conjunctiveness and hubness in the hippocampus.(A) Representational similarity analysis (RSA) logic. Left: associative similarity contrast, with expectedhigh regional representational similarity for comparisons of the same association, and low similarity forcomparisons of different associations, yielding a conjunctiveness metric for each voxel. Specific com-parisons were excluded to penalize perceptually driven effects (striped/blank cells): within-associationcomparisons with identical cue or associate stimulus categories (top left quadrant in matrix), andbetween-association comparisons with different cue and associate stimulus categories (bottom rightquadrant). Right: full condition-by-condition RSA contrast matrix used in the whole-brain searchlightapproach. Each cell represents a specific comparison between two conditions. Darkness indicatesdegree of expected pattern similarity. Four example comparisons are outlined. (B) Logic of networkanalysis. Whole-brain beta time-series correlation (left, five example voxel time series) was performedto obtain a voxel-by-voxel functional connectivity matrix (middle, darker shades indicate higher corre-lation coefficients). The participation coefficient was computed to obtain a hubness metric for eachvoxel, reflected by node size in the example graph (right, thickness of the edge relates to connectiv-ity strength). (C) Both RSA and network analysis show significant effects in the hippocampus (p <0.05 small-volume-corrected, thresholded at P< 0.05 uncorrected for display purposes) and overlap ofboth effects. (D) Hippocampal voxels showing overlapping effects were selected (left) to extract nor-malized similarity estimates (middle) for each comparison shown in a and hubness scores for ITI andrecall periods (right). *p < 0.05, **p < 0.005. Note that comparisons between these bars are shown fordisplay purposes only and reflect the effect shown in C for the selected hippocampal overlap voxels.(E) Observed Dice coefficient and relative overlap size (proportion of voxels showing overlap) of thehippocampus and associated p-value based on the null-distribution from the label shuffling (spatial re-sampling) procedure. Histogram y axis depicts the probability of observing a certain overlap statisticin randomly selected ROI (prob) on a logarithmic scale. The hippocampus shows significantly moreoverlap of conjunctiveness and hubness metrics than other regions in the brain.

Overlap between convergence metricsFinally, to assess the overlap of the hubness and conjunctiveness metrics, we thresh-olded and binarized both the conjunctiveness and hubness maps, and calculatedtheir intersection. As predicted, we observed overlapping patches of conjunctive-ness and hubness in the hippocampus (Figure 2.2C right panel, Figure 2.3C). Next,we defined the set of hippocampal voxels showing overlap on the group level asa region-of-interest (ROI) for post hoc analyses (Figure 2.2D). As expected, voxelsfrom the overlap ROI showed effects for both conjunctiveness and hubness metrics(see Figure 2.7 for an exploratory whole-brain connectivity analysis with the overlapROI as seed region). Moreover, we observed no associative similarity effect for thetemporal order in which an association was recalled (p > 0.26), but recall of the sameassociation was always more similar than recall of a different association, suggest-ing that the measured conjunctive representations are independent of the temporalorder of the stimulus pairs. In addition, we found no evidence for dependence ofthe associative similarity effect on the type of probe stimulus (p > 0.25, Figure 2.8,see section 2.4 for more details). To further investigate the relationship betweenconjunctiveness and hubness metrics in the overlap ROI, we performed an across-voxel correlation analysis within each participant (see section 2.4 for details). Onthe group-level average, voxels from the overlap ROI showed significant above-zero

2

34

Table 2.1: List of brain regions representing conjunctive information and their hubness scores.Inf, inferior; mid, middle; tri, triangular. Table denotes clusters with a minimal extent of 30 voxelsfrom the whole-brain conjunctiveness map, sorted on conjunctiveness peak T-value and thresholdedat p < 0.05 (nonparametric). Peak values for conjunctiveness (C, see Figure 2.3A) and hubness (H, seeFigure 2.3B) are displayed with their coordinates in Montreal Neurological Institute space. Nearestregion labels were obtained using the AAL atlas. Statistics of the two hippocampal peak locationsreported in the main text (one for conjunctiveness in left hippocampus and one for hubness in righthippocampus, see Figure 2.2C) are denoted at the bottom for comparison.

Anatomical region x y z T-value C T-value H

right supramarginal 64 -40 40 4.63 -0.05right frontal inf tri 54 28 30 4.38 0.57left angular -44 -58 32 4.37 -0.28left precuneus -8 -62 44 3.83 -1.01left temporal mid -52 -40 -8 3.74 1.6left hippocampus -32 -16 -8 3.64 0.93left supp motor area -12 -10 54 3.31 -0.52left cerebellum 6 -12 -64 -28 2.79 0.82right thalamus 12 -12 4 2.47 0.73right hippocampus 30 -16 -14 2.32 2.77right temporal mid 54 -18 -10 2.25 -0.98left temporal mid -58 6 -30 2.22 0left temporal inf -64 -58 -8 2.08 0.66right temporal inf 36 4 -48 1.54 0.23right precuneus 8 -76 60 1.51 1.09right cerebellum crus1 58 -64 -34 0.48 0.46left precentral -40 -24 72 0.35 0.41

Hippocampal region-of-interestleft hippocampus (C-peak) -30 -16 -14 3.59 3.07right hippocampus (H-peak) 28 -14 -22 1.71 3.75

correlation coefficients (Wilcoxon signed-rank test: Z = 2.21, p = 0.026), indicatinga relationship between conjunctiveness and hubness metrics. But, how surprisedshould one be to observe overlap specifically in hippocampus? To answer this ques-tion, we performed a ROI-based spatial resampling procedure, designed to assesswhether the observed overlap was greater than potential spurious overlap at a cer-tain threshold and quantify this expression in a p-value (see section 2.4 for details).In this analysis, we calculated two complementary overlap statistics, namely Diceoverlap coefficient and relative overlap size, using all voxels from the left and righthippocampus (Figure 2.2C) (Dice, 1945), as defined by the Automated AnatomicalLabeling atlas for SPM8 ((Tzourio-Mazoyer, 2002)). We then compared the resultingoverlap metrics to a null-distribution obtained by resampling with randomly per-muted region labels using all 116 atlas regions. In other words, for each permuta-

2

35

tion, we computed the overlap scores for voxels from two randomly selected re-gions, which yielded our null-distribution. The hippocampus showed significantlymore overlap of representation and connectivity metrics (Dice coefficient: 0.25, p =0.0135; relative overlap size: 5%, p = 0.0004) than expected by chance (Figure 2.2E).These results were robust to various cutoff values used to threshold the two maps(Figure 2.9). Notably, we did not observe this effect when we substituted either theconjunctiveness or hubnessmapwith a univariate activity mapwhere we contrastedthe retrieval phase with the inter-trial intervals (univariate with conjunctiveness orhubness: for all combinations Dice coefficient: 0, p = 1; relative overlap size: 0%, p= 1, see Figure 2.10 for whole-brain univariate results). These results suggest thatthe significant overlap of conjunctiveness and hubness in the hippocampus are notexplained by univariate signal differences.

2.3.DiscussionUsing a combined approach of representational similarity and network analyses, weprovide evidence for mnemonic convergence in the human hippocampus. Our find-ings highlight the key role of the hippocampus in representing conjunctive informa-tion and relate this function to its importance in connecting subnetworks duringmemory retrieval. We demonstrate that this crucial role of the hippocampus as aconnector hub is notably prevalent during memory retrieval, at the same time whenconjunctive representations are reactivated (Staresina et al., 2012).The present results are in line with both theoretical work and empirical findings. Hip-pocampal place cells integrate the spatial features characterizing a specific locationand have been put forward as the essential elements of a map-like representationof the environment (O’Keefe and Nadel, 1978). In addition, other types of high-levelconjunctive cells have been observed in the hippocampal formation, such as cellscoding for conditioned behavioural responses (Eichenbaum et al., 1989), specific ol-factory cues (Berger et al., 1976), the conjunction of location and heading directionof an animal (Sargolini, 2006), or location in conjunction with a remembered object(Moita et al., 2003). These conjunctive representations constitute the hallmark ofepisodic memory, as they represent the relations between elements of an episode(Eichenbaum, 2000; Wood et al., 1999). Our results are consistent with the idea thatthe hippocampus contains these index-like representations (Teyler and DiScenna,1986) in sparse networks (Quiroga et al., 2005; Wixted, 2014), binding multiple corti-cal elements of an episode intomemory (Davachi, 2006; Horner et al., 2015; Hannulaand Ranganath, 2008; Nadel and Peterson, 2013; Eichenbaumet al., 2007; Horner andBurgess, 2014).Evidence from animal electrophysiology as well as human lesion and anatomicalconnectivity studies posits that the hippocampus acts as a major network hub dur-ing retrieval: the hippocampus ultimately receives input from most regions of the

2

36

brain via the entorhinal cortex and thus is an anatomical hub (Lavenex and Ama-ral, 2000; Bota et al., 2015). The hippocampus constitutes the apex of the visualprocessing hierarchy since it receives converging inputs from most upstream visualregions (Felleman and Van Essen, 1991). Graph-theoretical analyses of human diffu-sion tensor-imaging data have revealed that the hippocampus is part of a so-calledrich-club of network hubs, characterized by denser connectivity among club mem-bers than with less connected regions (van den Heuvel and Sporns, 2011; Misic et al.,2014). This finding is well in line with the results of our network analysis, where weidentify the hippocampus as a connector hub during memory retrieval. By using theparticipation coefficient to quantify hubness from whole-brain connectivity data inour graph-theoretical analysis, we summarize the importance of the hippocampusfor interactions between distributed subnetworks. Moreover, by contrasting thishubness metric during retrieval against the inter-trial intervals, we are able to iso-late task-related contributions. Note, however, that we used the inter-trial intervalas baseline, and therefore the observed relative participation coefficient increasemight be due to a hubness decrease during the inter-trial interval. Nevertheless, theobserved increase in participation coefficient suggests a more prominent role forthe hippocampus during the retrieval of memories, likely represented in distributedparts of the brain. We showed that the hub status of the hippocampus is linked tothe memory retrieval phase in our task, which accords with the large body of hu-man neuroimaging evidence implicating the hippocampus in memory retrieval (Zei-thamova et al., 2012a; Ranganath et al., 2004), as well as recent electrophysiologicalstudies suggesting that the hippocampus serves as a network communication hubfor memory (Watrous et al., 2013; Battaglia et al., 2011).Moreover, we provide experimental evidence that these network characteristics ofthe hippocampus directly relate to its representational role: conjunctive coding ofassociative information. We observe both hub-like properties and conjunctive rep-resentations in the hippocampus, suggesting that the hippocampus acts as a con-vergence zone. This notion fits with computational models and general principlesof brain function (Buckner and Krienen, 2013), which recognize convergence as a keymotif in the brain (Papez, 1944). Information about the external world is processedby sensory regions and progressively integrated as it reaches upstream brain areasand is ultimately evaluated by decision-making systems (Shohamy and Turk-Browne,2013). By demonstrating the hub role of the hippocampus during memory retrieval,we provide a strong link between functional connectivity and functional specializa-tion formemory processes: as an important connector hub, the hippocampus is ableto integrate information frommultiple subnetworks into a coherent conjunctive rep-resentation, consistent with the convergence motif.Althoughwe did not aim to investigate functional specialization along the hippocam-pal long-axis, we observe overlap between representation and connectivity in the

2

37

middle and anterior part of the hippocampus. There is substantial evidence for afunctional specialization along the posterior-anterior axis of the hippocampus (Pop-penk et al., 2013; Ranganath and Ritchey, 2012; Collin et al., 2015). Our findings mayrelate to the preference of the anterior hippocampus for non-spatial stimulus mate-rial or more abstract, higher-level associative information (Collin et al., 2015), such astemporal order-invariant conjunctions relevant in the current experiment (Gutchessand Schacter, 2012).In conclusion, we show that the human hippocampus acts as a mnemonic conver-gence zone, characterized by both hub-like network connectivity and conjunctiverepresentations. We thereby provide evidence for the long-held hypothesis thatthe hippocampus binds distributed information into memories. Furthermore, weoutline a quantitative method to investigate convergence zones in humans, whoseexistence has been hypothesized for a long time by computational models. Futureapplications of our approach could leverage thismethod to track the dynamics of hip-pocampal processing during memory consolidation and to investigate the integrityof the hippocampus during normal or pathological ageing.

2.4.Methods2.4.1. ParticipantsThirty-five participants (19 females, average age: 22.7 years, range: 18-32 years) took part inthe study. All were in good health, with no history of psychiatric or neurological diseases,no brain abnormalities and normal or corrected-to-normal vision. Before the experiment,participants gave their informed consent and were reimbursed for their participation. All ex-perimental procedures were approved by the local ethical review committee (CMO regionArnhem-Nijmegen, TheNetherlands). Five participants were excluded due to technical prob-lems with the scanner and an additional five participants since they were unable to reach asufficient performance level (d-prime < 1.0). Therefore, the data of 25 participants (15 females,average age: 22.7 years, range: 18-32 years) entered our analysis. The sample size was basedon previous RSA studies on memory and the hippocampus (Milivojevic et al., 2015; Collinet al., 2015).

2.4.2. StimuliWe used grayscale images of faces (Karolinska Directed Emotional Faces) (Lundqvist et al.,1998), houses (Stanford Vision Lab stimulus set) (Oliva and Torralba, 2001) and human bodies(Bodily Expressive Action Stimulus Test set) (de Gelder and Van den Stock, 2011). All imageswere cropped to 200 × 200 pixel dimensions and normalized using the SHINE toolbox forMATLAB (Willenbockel, 2010) (v2014a, The MathWorks) by adjusting the mean luminanceand s.d. of the intensity values for each pixel. Stimuli were presented to participants usingthe Presentation software package (v16.4, Neurobehavioral Systems).

2

38

2.4.3. Paired-associate learning before the scanning experimentParticipants commenced with the initial encoding session outside the scanner, separatedinto six study and test cycles. During study cycles, participants learned 12 random associa-tions between pairs of pictures. Associations comprised face-house, face-body and house-body pairs (four pairs of each type). In a study block, the 12 pairs were presented in randomorder. In each trial, the two stimuli of each pair were shown in succession (1,000 ms on-screen, 1,000 ms inter-stimulus interval). We used an inter-trial interval of 3,000 ms, duringwhich a fixation dot was presented. Order of presentation of the two stimuli per pair wascounterbalanced across cycles. In the test blocks, 48 test trials were presented in which oneof the stimuli of each pair was presented as a retrieval cue, followed by a probe stimulus,which could either be the associate (match probe) or a different stimulus from the same cat-egory (non-match probe). Either of the pair members could appear as a cue, with the ordercounterbalanced within and across cycles. The cue and probe stimuli were each presentedfor 200ms. Cue and probe presentations were separated by a retrieval phase of 1,000, 3,000or 5,000 ms (counterbalanced across cues, pairs, matching probe and cycles) during whichparticipants were asked to retrieve the specific associate of the cue. Participants were in-structed to respond as fast as possible with their right hand, using two response buttons,and to indicate whether the probe matched the associate (hit or false alarm) or not (correctrejection or miss). Response mapping of these two buttons was counterbalanced across par-ticipants. The maximal response window was set to 600 ms. If participants did not respondwithin the response window, then a too-late message was presented for 1,000 ms. The vari-able retrieval phase together with the short response window ensured that participants hadto respond promptly to elicit immediate memory retrieval. After each response, feedbackwas provided by presenting the associate (1,000ms on screen). Trials were separated by vari-able inter-trial intervals of 1,000, 3,000 or 5,000 ms (retrieval phase and inter-trial intervaladded up to 6,000 ms in each trial). During a given test block, each association was tested4 times. At the end of each test block, the percentage of correctly responses was displayedto the participant. We encouraged participants, by way of a monetary reward (a bonus of 5Euros), to reach a minimum of 80% correct responses (hits and correct rejections) in at leastone of the test blocks, in order to foster high memory performance.

2.4.4. Retrieval task in the scannerAfter a 30-min break, participants performed the retrieval task in theMRI scanner in 2 runs ofapproximately 25 min each, with a short half-time break in-between lasting approximately5 min. During the scan session, a total of 288 retrieval test trials were presented to theparticipant (144 trials in each run). Trial structure was identical to the combined test blocksof the encoding session. However, we did not provide feedback and set the retrieval phaseand inter-trial interval lengths to 1,000, 6,000 and 11,000 ms, respectively. The performancescore was only displayed at the end of the experiment. The pairs were presented 12 timesin each run: 6 times for each of the two possible temporal cue-associate orders. Conditions,trial durations and match probes were counterbalanced within each run. Trial order wasrandomized in both runs.

2

39

2.4.5.Data acquisitionNeuroimaging data were acquired using a 3-T MR scanner (TIM Trio; Siemens Healthcare)in combination with a 32-channel head coil. For the functional scans, we used a three-dmensional (3D) echo planar imaging (EPI) sequence (voxel size: 2 × 2 × 2 mm, volumeTR: 1,800 ms, TE: 25 ms, flip angle: 15 degrees, 64 slices, FOV: 224 × 224, orientation: -25 de-grees from transverse plane, GRAPPA acceleration factor: 2, acceleration factor 3D: 2) (Poseret al., 2010). Using the AutoAlign head software by Siemens, we ensured a similar FOV tiltacross participants. Functional scan runs contained between 1032 and 1093 volumes, sincethe instruction screens were self-paced. In addition, we acquired field maps using a gradientecho sequence (voxel-size: 3.5 × 3.5 × 2 mm, volume TR: 1020 ms, TE1: 10.00 ms, TE2: 12.46ms, flip angle: 90 degrees, 64 slices, FOV: 224 × 224, orientation adjusted to functionalsequence, descending slice order). At the end of the scanning session, we obtained a struc-tural scan using an MPRAGE sequence (voxel-size: 1 × 1 × 1 mm, volume TR: 2,300 ms, TE:3.03 ms, flip angle: 8 degrees, FOV: 256 × 256, ascending slice order, GRAPPA accelerationfactor: 2, duration: 5:21 min).

2.4.6. fMRI preprocessingWe preprocessed MRI data using the Automatic Analysis framework, which combines toolsfrom SMP8, FreeSurfer v5.1, and the FMRIB Software Library v5.0, , complemented by cus-tom scripts. The preprocessing pipeline consisted of the following steps: we removed bi-ases resulting from field inhomogeneities from the native structural images using the SPM8new segment option. Furthermore, we denoised the structural images using an AdaptiveOptimized Nonlocal Means filter (MRI denoising software package) (Manjon et al., 2010).Next, we performed a premasking procedure to exclude the neck from the structural im-age using a template image and ran a Freesurfer brain extraction and SPM segmentationprocedure to obtain segmentation masks for grey matter, white matter, cerebrospinal fluidand out-of-brain voxels. Furthermore, we realigned and unwarped the functional imagesusing the fieldmap images. In addition, we employed a spike-detection algorithm to recordand later model signal spike events as nuisance variables. Functional and structural imageswere coregistered to a functional template (mean EPI) and a structural template respectively,after which the functional images were registered to structural space. We extracted thesignal time course from white matter, cerebrospinal fluid and out-of-brain voxels and in-cluded these as nuisance variables. Field bias was removed from the mean EPI after whichwe performed a Freesurfer brain extraction procedure to obtain a brain mask. To account forinter-subject differences in brain morphology, we constructed a group structural templateusing the Advanced Normalization Tools toolbox v1.9. Subsequently, we used the parame-ters obtained via this procedure to later normalize our single-subject statistical maps to anintermediate common space as a final step, before transforming toMNI space using FMRIB’sLinear Image Registration Tool and performing the group-level statistical analysis.

2.4.7.General linear modellingOurmain analyses were restricted to the retrieval phase in each trial and we included all 288trials in our analyses. We used all trials (trials with correct responses and the small numberof trials with incorrect responses) since we aimed to obtain the most reliable estimate ofresponse patterns, by constructing balanced regressors containing three trials of each con-

2

40

dition: wemodelled brain activity during the retrieval phase and inter-trial intervals by usingthree randomly selected trials with a short, medium and long duration from the same runand condition (144 trials, resulting in 48 regressors per functional run) using boxcar func-tions spanning the respective intervals. The three trials selected for a given condition weremaximally spaced apart in time. By modelling three trials with different retrieval phase du-rations from the same condition, with different onset spreads across the experiment, weaimed to minimize influence of time-dependent effects, such as temporal autocorrelationand habituation effects, and thereby obtain a more reliable set of beta estimates for each ex-perimental condition. We included the small number of incorrect response trials to be ableto balance the total amount of delay for each regressor (one short, one medium and onelong trial) and be able to utilize the short-delay trials, at the expense of making our analy-sis potentially more conservative. Inter-trial intervals were explicitly modelled to obtain thebeta estimates required for the network analysis. For each condition-specific regressor (con-taining the retrieval phases of three delay intervals), we ran a general linear model (GLM)including the regressor-of-interest and one single additional regressor containing all otherconditions and other task (that is, regressors for faces, scenes, bodies, probes, retrieval cuesand button presses) and nuisance variables (Mumford et al., 2012), using standard SPM func-tions with default settings. Both runs were modelled together in each GLM, accounting forgeneral differences between the runs. In total, we obtained the beta images for 96 retrievalphase regressors and another 96 complementary inter-trial interval regressors (48 per func-tional run). Decorrelating regressors for different groups of trials from the same conditionusing this iterative method yields beta weights well-suited for multivariate pattern analysison event-related designs (Mumford et al., 2012).

2.4.8. Searchlight representational similarity analysisWe performed a whole-brain searchlight analysis to assess which regions contained mul-tivoxel information about specific memory representations (Kriegeskorte et al., 2006). Af-ter applying a grey matter mask, we extracted the multivoxel activity pattern within eachspherical searchlight (4 voxel radius, including a minimum of 30 grey matter voxels), fromeach of the 96 retrieval phase beta images. Similarity between patterns was computed us-ing Spearman’s correlation to account for nonlinear effects and deal with outliers withoutspecifying an arbitrary threshold (Zheng, 2013). We then constructed a balanced regressor-by-regressor contrast matrix for the hypothesized representational similarity pattern, witha mean value of 0. The observed similarity space of each sphere was then fitted to the con-trast matrix, using a GLM. The resulting parameter estimates were assigned to the centrevoxels of each sphere. To correct potential biases in the T-value distributions and to equalizevariance across participants, we applied a mixture model to our T-maps. We then warpedthe resulting statistical maps to MNI space and performed additional smoothing (full-widthat half maximum (FWHM): 2 mm) to improve spatial alignment across participants.

2.4.9. Conjunctive mnemonic information contrastTo be sensitive to conjunctivememory retrieval in our analysis, we defined a specific contrastwhere we expected high pattern similarity when comparing the multivoxel activity patternsof a specific association to a different instance of the same association (associative similar-ity contrast, Figure 2.2A). Conversely, when we compared the patterns during retrieval of a

2

41

specific association with the pattern in response to a different association, we expected highdissimilarity. To control for unspecific perceptual effects and to maximize our sensitivity formnemonic representations, we introduced a perception penalty by excluding specific com-parisons: whenever we compared neural patterns of two instances of the same association,the cue-associate order of one of the instances was always reversed. Conversely, when wecompared instances of different associations, we made sure that cue-associate order wasidentical. Any perceptual similarity effects driven by the visual categories of the cue andassociate were thus minimized.

2.4.10. Functional connectivity analysisFor the connectivity analysis, we concatenated beta estimates for regressors of the retrievalphases (used for the RSA) and inter-trial interval separately, resulting in two beta vectorsper voxel. After spatial subsampling (resulting in a voxel size of 8 × 8 × 8 mm) we com-puted voxel-wise spatial correlation coefficients of the beta vectors to quantify functionalconnectivity for each condition. All following analyses were performed on the weightedconnectivity matrices, where negative correlations were set to zero (Power et al., 2013) andall positive edges were thresholded at p < 0.05 (false discovery rate corrected), to preservesignificant connections. We indexed hubness by estimating the participation coefficient,quantifying the distribution of voxel-wise connections among local subnetworks. To assigneach voxel to a subnetwork, we derived an additional 116 × 116 region-by-region connectiv-ity matrix from the averaged beta vectors, where regions were defined using the AutomatedAnatomical Labeling (AAL) atlas (Tzourio-Mazoyer, 2002). Subsequently, after thresholding(edges > 0, p < 0.05 false discovery rate-corrected) we parcellated the 116-node network us-ing modularity detection (Louvain method (Blondel et al., 2008)) and assigned each voxel toone of the resulting subnetworks. We computed the participation coefficient for eachvoxel by closely following the procedure employed by Power and colleagues (Power et al.,2013). is given by:

∑(̂)

Here ̂ is is the number of edges of voxel to voxels in subnetwork , while is the totalamount of connections of voxel , and is the number of subnetworks. This procedureresulted in a normalized voxel-wisemeasure ranging from0 (provincial hub: only connectingwithin subnetwork) to 1 (connector hub: only connecting between subnetworks). Next, wetransformed the hubness maps to MNI space and contrasted the retrieval phase with theinter-trial intervals (Figure 2.3B).

2.4.11. Statistical analysis of conjunctiveness and hubness mapsTo test whether voxels in the hippocampus show significant effects, we used FSL RAN-DOMISE to obtain nonparametric statistics with 10,000 random permutations. The teststatistic was based on a one-sided t-test of within-subject difference maps, with 5-mm vari-ance smoothing and threshold-free cluster enhancement (Smith and Nichols, 2009). Wecorrected for multiple comparisons using FWE correction, restricted to a small-volume com-prising bilateral hippocampus, as defined by the AAL atlas. All whole-brain maps presented

2

42

in the current work were thresholded with voxel-wise nonparametric p-values obtained us-ing FSL RANDOMISE. We obtained post hoc modelled mean pattern similarity estimates forthe separate comparisons (for example, same association with same order, same associationwith different order, different association with same order, and different association with dif-ferent order) using four contrasts of the isolated comparisons against zero. We fitted thesecontrasts using a GLM and averaged the beta estimates, reflecting neural similarity, fromall hippocampal voxels showing overlap of the RSA and functional connectivity analysis (involume space, see the overlap ROI, Figure 2.2D). The magnitude of these beta estimates wasthen normalized by demeaning across the four conditions within each participant. In addi-tion, we extracted the hubness estimates from the overlap ROI for the ITI and recall periods.Comparisons between these measures (p-values obtained using two-tailed nonparametricpaired t-tests with 100.000 permutations) were added for display purposes (Figure 2.2D). Toinvestigate the relationship between hubness and conjunctiveness metrics, we computedSpearman’s correlation coefficients across voxels from the overlap ROI. Spearman’s coeffi-cients were used to account for nonlinear effects. Next, we tested for a significant positiveor negative relationship on the group level, using a two-tailed Wilcoxon signed-rank test.For visualization of the imaging results, whole-brain cortical and cerebellar surfaces render-ings were created using the brain visualization tool CARET v5.65. Note that these surfacerenderings were only used for visualization. All statistical tests were performed on the vol-ume maps. To illustrate the main effect in the hippocampus, volume maps are shown inFigure 2.2C.

2.4.12. Statistical analysis of overlap between convergence metricsTo test regional coincidence of hubness and conjunctiveness, we opted for a hypothesis-driven, yet full-brain resampling approach: first, we defined our predicted bilateral hippocam-pal ROI as the corresponding anatomical masks extracted from the AAL atlas. Next, wecomputed summary overlap statistics for our anatomical hippocampal ROI. We binarizedour voxel-wise network centrality map in MNI space, yielding a binary vector defining so-called hub voxels for our anatomical ROI. This procedure was repeated for the conjunctiveinformationmap to obtain a binary vector defining the informative voxels in the hippocam-pus. Both the conjunctiveness and hubness maps were thresholded at p < 0.05 uncorrected,using the voxel-wise nonparametric p-values. We computed two complementary metricsto quantify overlap: first, we used the Dice coefficient to assess the specificity of overlapbetween hubness and conjunctiveness effects, regardless of extent and region size. Heredouble the length of the logical conjunction between and is divided by their summedindividual lengths (that is, the sum of all logical true elements in both vectors separately):

| ∩ || | | |

Second, to quantify the extent of overlap, relative to the total region size, we computed pro-portion of voxels that show both hubness and conjunctiveness effects of our hippocampalROI containing a total number of voxels , using the following equation:

| ∩ |

2

43

This procedure yielded two complementary overlap measures for the left and right hip-pocampus, on which we subsequently performed a spatial permutation test. Here we com-puted the same overlap score for randomly selected ROIs with 10,000 permutations (tworandom subregions from the AAL atlas in each permutation). For the crucial final statis-tical test, we hypothesized that no 45% of all of these random ROIs would yield overlapscores higher than the overlap scores observed in the bilateral hippocampus. We investi-gated whether the cutoff nonparametric p-value used to threshold the input vectors and

influenced the results. To this end, we repeated the procedure and plotted the corre-sponding probability of observing a higher overlap score in a random ROI as a function ofthe critical p-value used to threshold the input vectors (Figure 2.9).

2.4.13.Univariate activity contrastTo test whether effects resulting from hubness or conjunctiveness metrics could be ex-plained by univariate effects, we smoothed our data (FWHM: 8 mm) and applied a GLMincluding regressors for retrieval phases, inter-trial intervals, faces, scenes, bodies, probes,retrieval cues and button presses for each functional run. Next, we contrasted the betaimages of the retrieval phases with the beta images of the inter-trial intervals (Figure 2.10).Univariate activity maps were analysed in the same way as the conjunctiveness and hub-ness maps, that is, warped to MNI space via Advanced Normalization Tools common space,before obtaining nonparametric statistics.

2.4.14.Head displacement analysisTo rule out potential head movement biases in our network analysis (Ekman et al., 2012), wecompared the root-mean-square of all six translation parameters of the retrieval and inter-trial intervals. A t-test revealed no significant differences between conditions (T₂₄ = 0.05, p> 0.96). In addition, a histogram of mean displacements magnitudes revealed no apparentdifferences on a finer scale (Figure 2.4).

2.4.15. Eigenvector centrality analysisTo corroborate our participation coefficient results and evaluate the robustness of our con-nectivity findings, we repeated our analysis with a different centrality measure and alterna-tive preprocessing. Here we followed the procedures used by Ekman et al. (Ekman et al.,2012): we extracted coregistered time series from all grey matter voxels and shifted the timecourse by 3 volumes (5.4 s) to compensate for the hemodynamic response lag. We regressedout head motion and out-of-brain signal from the time series, followed by a spatial subsam-pling procedure, resulting in a voxel size of 4 × 4 × 4 mm. Next, we computed voxel-wisespatial correlation coefficients of the retrieval phases and inter-trial intervals separately. Allsubsequent analyses were performed on the weighted connectivity matrix, where negativecorrelations were set to zero. We derived a centrality score for each individual voxel by com-puting the eigenvector of the connectivity matrix with the highest eigenvalue. Comparedwith the participation coefficient, eigenvector centrality is a coarser hub measure, that indi-cates how important (that is, central) regions are within the global network. We followed aprocedure similar to the participation coefficient analysis, where we transformed the eigen-vector centrality maps to MNI space and contrasted retrieval phase with the inter-trial inter-vals (Figure 2.5).

2

44

2.4.16. Seed-based connectivity analysisFor the exploratory seed-based connectivity analysis, we used the same recall and ITI betatime-series constructed for the network analysis. We back-warped the ROI mask with thehippocampal overlap voxels (Figure 2.2D) to individual participant brain space. After apply-ing spatial smoothing (FWHM: 8 mm), we extracted the mean time course of the overlapROI and computed spatial correlation coefficients with all brain voxels. Coefficients of therecall and ITI phases were warped to MNI space, Fisher’s Z-transformed and contrasted, toobtain a normalized whole brain difference map (Figure 2.7).

2.4.17. Probe type control analysisAlthough we excluded the probe presentation interval from our recall regressors and explic-itly modelled probe stimuli as nuisance in our initial GLM, it is important to investigate theinfluence of probe type: when comparing twomatching-probe trials of the same association,but with different order, participants ultimately view the same two stimuli, whereas in thenon-match probe trials only the cue stimulus is shared. Therefore, as we argue that our RSAis sensitive to mnemonic representations, the associational similarity effect should not bepredominantly driven by the match probe trials. To assess whether our associative similarityeffect in the hippocampal overlap ROI is driven by probe type, we performed an additionalGLM analysis with separate regressors for match and non-match probe trials (Figure 2.8).The obtained similarity estimates for the two main comparisons of interest (that is, same as-sociation with different order, different association with same order, see Figure 2.2A) weredemeaned and contrasted (P-values obtained using two-tailed nonparametric paired t-testswith 100,000 permutations). Note that these match and non-match contrasts are less sensi-tive, since they are based on half the amount of comparisons entering the main associativesimilarity contrast.

2

45

2.5. Supplemental Information2.5.1. Supplemental Figures

40T

B

A

C

R

A

V

Figure 2.3: Distribution of conjunctivemnemonic information, increased hubness and overlapacross the brain.(A) Whole-brain conjunctiveness scores from the RSA, rendered on a cortical surface map. These sur-face renderings were only used for visualization; the statistical tests were carried out on volume maps.Regions with a high T-value exhibit higher neural pattern similarity when comparing instances of thesame association relative to comparing different associations (associative similarity, see Figure 2.2A).In addition to the hippocampus, evidence for conjunctive codingwas observed in an extended networkof regions that have been previously implicated inmemory processes, including prefrontal cortex (Ran-ganath et al., 2004; Henson et al., 1999; Nolde et al., 1998; Rugg et al., 1999; Wagner et al., 2001; Buckner,1996; Miller, 2000; Euston et al., 2012), lateral parietal cortex (Wagner et al., 2005), precuneus (Fletcheret al., 1995) and lateral temporal cortex (Binder and Desai, 2011; Patterson et al., 2007). Note the ab-sence of high values in early visual regions, indicating that the RSA is not sensitive to visual categoryeffects. (B) Whole-brain participation coefficient hubness scores. Regions with a high T-statistic areprominent connector hubs duringmemory retrieval and therefore relatively important for interactionsbetween subnetworks. The reverse contrast (inter-trial interval versus retrieval condition participationcoefficients) yielded no significant differences (p > 0.79 whole-brain FWE- corrected) and no clusterswith a minimal extent of 30 voxels thresholded at p < 0.05. Maps in (A) and (B) thresholded at p < 0.05to illustrate which voxels were used to compute the overlap scores for the spatial permutation test(see Figure 2.2E). Circles indicate p < 0.05 small-volume FWE-corrected peaks in the hippocampus. (C)Binary overlap between hubness and conjunctiveness maps, obtained by intersecting (A) and (B). Notethat overlap is exclusive to the bilateral hippocampus.

2

46

fre

qu

en

cy (

%)

0 1 2 3 4 50

10

head motion across all participations (mm)

20

30retrieval phase

inter-trial phase

Figure 2.4: Head displacement during retrieval and inter-trial intervals.Histogram of average head movements across participants derived from the realignment parameters.No significant difference was observed between conditions (T₂₄ = 0.05; p = 0.96, see section 2.4 fordetails).

0.84

0.73

PC

Figure 2.5: Participation coefficient scores during inter-trial intervals.We observed regions previously associated with high participation coefficient (PC) scores during restblocks, such as lateral temporal gyrus, posterior cingulate, fusiform gyrus, insula and inferior frontalgyrus (Power et al., 2013). Map was thresholded by setting all voxels below 99% of robust range (ofnon-zero voxels) to zero.

5

0

T

Figure 2.6: Whole-brain eigenvector centrality scores.Recall versus inter-trial interval differencemap computed from raw time series data (see section 2.4 fordetails). Effects are similar to the participation coefficient results (see Figure 2.3). Map thresholded at p< 0.05 for comparison. Circle indicates p < 0.05 small-volume FWE- corrected peak in the hippocampus.

2

47

5

0

T

Figure 2.7: Seed-based connectivity profile of hippocampal overlap voxels.To investigate the functional connectivity of the hippocampus with other brain regions during retrieval,we performed an exploratory seed-based connectivity analysis, with the voxels in the hippocampusthat showed effects for both hubness and conjunctiveness (see Figure 2.2C, right panel) as seed ROI.Map shows brain regions that potentially drive the observed participation coefficient increase duringmemory retrieval. We observed no significant differences (p > 0.30 whole-brain FWE-corrected). How-ever, inspection of the seed-based connectivity map at a more liberal threshold (p < 0.05) revealedincreases of hippocampal connectivity with an extended network of task-related brain regions, includ-ing ventromedial prefrontal cortex (Euston et al., 2012), sensorimotor cortices and parietal associationareas (Wagner et al., 2005; Hoesen et al., 1972). Note that due to this liberal threshold, the results fromthis analysis should be interpreted with caution.

ne

ura

l sim

ila

rity

(a

.u.)

match non-match

probe type

Figure 2.8: Associative similarity effect split for match and non-match probe types.Results from an additional exploratory GLM analysis to assess associative similarity for trials with aprobe that matched the associate stimulus separately from trials with a probe that was a differentstimulus from the same category. Bars are displaying estimates from the hippocampal overlap voxels(see Figure 2.2D). Note that these comparisons (dark purple: association with different order, lightpurple: different association with same order, see Figure 2.2A) are based on half the amount of trialscompared to the associative similarity effect presented in the main text. The non-match conditionappears consistent with the associative similarity effect, although we observed no significant differ-ences in either of the probe type conditions (p > 0.25). The apparent (but non-significant) reversal ofthe effect for match trials with respect to the main analysis, in which match and non-match trials arecombined, is likely due to the differing variance of the match and the non-match trials.

2

48

10-1 10-2100

10-4

threshold

(p-value)

ran

do

m D

ice

ov

erl

ap

(pro

ba

bli

ty)

p < 0.05

10-1100

10-4

threshold

(p-value)

ran

do

m o

ve

rla

p s

ize

(pro

ba

bil

ity

)

p < 0.05

BA

Figure 2.9: Overlap at various thresholds.The impact of different thresholds applied to the conjunctiveness and hubness maps on the specificityof the effect in the hippocampus for the overlap statistics. Logarithmic Y-axis denotes the probabilityof observing a higher overlap score in random regions-of-interest at a given threshold of the brainmaps,plotted on the logarithmic X-axis. Values above the gray dotted line indicate (A) significantlymoreDiceor (B) relative overlap of conjunctiveness and hubness metrics in the hippocampus. The hippocampusgenerally shows significantly more overlap than random brain regions, unless the applied thresholdbecomes too conservative (p < 0.02). Gray arrows indicate the threshold used in the main analysis (p< 0.05).

9

0

T

Figure 2.10: Brain regions showing more activation during memory retrieval compared to theinter-trial intervals.Results from a univariate GLM analysis contrasting retrieval versus inter-trial interval activity. Mapthresholded at p < 0.05 to illustrate which voxels contributed to the overlap scores. Note that thehippocampus does not show increased univariate activity during retrieval, evenwhen adopting a liberalthreshold.

3Increased hippocampal neural patternsimilarity of newly associated stimuli

Alexander R. Backus, Lea Himmer, Christian F. Doeller

This chapter is in preparation as: Backus, A.R., Himmer, L., Doeller, C.F., Increased hippocampal neuralpattern similarity of newly associated stimuli.

51

3

52

3.1. IntroductionDuring everyday life, we continuously acquire new knowledge through associativelearning. From the simple association between an object’s shape and color, to themyriad of associations between locations, sights, sounds, smells, people and ab-stract concepts that define rich episodic memories (Tulving et al., 1972). Together,these memories form an interconnected malleable network of associations, that isperpetually updated and expanded with new information, and ultimately gives riseto our knowledge base (Milivojevic and Doeller, 2013). But how are these memorynetworks shaped and where are they represented in the brain?Computational models and experimental evidence suggest that the hippocampusis the critical brain region involved in the formation of associative memories (Marr,1971; Eichenbaum et al., 2007; Burgess, 2002; Stark and Squire, 2001; Davachi, 2006).Through its dense, hub-like connectivity with other brain regions (Chapter 2) andunique conjunctive representations (O’Keefe and Nadel, 1978; Eichenbaum et al.,1999), the hippocampus has been attributed a binding role, functioning as a con-vergence zone for disparate sources of information in the brain(Damasio, 1989). Ac-cordingly, hippocampal neurons have shown to adapt their stimulus-specific firingpatterns as a function of learning (Cahusac et al., 1993;Wirth, 2003;Wirth et al., 2009;Reddy et al., 2015; Ison et al., 2015; Sakai and Miyashita, 1991). In addition, neuralstimulus representations, expressed in the patterns across multiple neurons, havebeen found to become more distinguished over the course of learning (McKenzieet al., 2013, 2014). But can we measure the formation of such associative memorynetworks in humans, using non-invasive neuroimaging techniques?Recent neuroimaging studies have investigated reconfigurations of neural represen-tations as a function of learning, by using a combination of functional magnetic res-onance imaging (fMRI) and representational similarity analysis (RSA) (Kriegeskorteet al., 2006). In these studies, changes in the similarity structure of regional responsepatterns was investigated, from before to after learning. Using this technique (here-after referred to as differential RSA), representational network reconfigurations havebeen demonstrated in fear conditioning (Visser et al., 2011), implicit temporal regular-ity learning (Schapiro et al., 2012, 2013, 2015), transitive inference (Schlichting et al.,2015) and narrative insight (Milivojevic et al., 2015; Collin et al., 2015). However, itremains untested whether a representational change can also be detected in thecase of explicitly paired associates. Investigating this open issue is of particular im-portance, since paired-associate learning is a simple and well-understood paradigmto model episodic memory in humans (Calkins, 1894) close to the single-cell record-ing experiments in animals and humans (Ison et al., 2015; Sakai and Miyashita, 1991).Moreover, previous work has focused on relatively small, albeit significant, represen-tational reconfigurations across groups of individuals. The possibility to interrogateindividual memory networks using the differential RSA remains unaddressed.

3

53

Here, we present the litmus test for the differential RSAmethod. We used an explicitpaired-associate learning task with fractal-like visual stimuli, in which we aimed toovertrain four simple random associations (Figure 3.1). We acquired fMRI data beforeand after learning, allowing us to quantify the representational change due to asso-ciative learning and probe individual memory networks. We hypothesized that theprime target brain region to represent these newly learned associative memory net-works would be the hippocampus. In addition, we further explored the distributionof representational reconfigurations across the entire brain. Finally, we systemati-cally investigated predictive value of individual neural similarity data, to gauge thepotential of differential RSA for reading-out neural memory networks and trackingknowledge acquisition.

memory testspre learn post

1-5s

1s

1-5s

1s

C

A

B

D

2s

8s

2s

8s

...

pre learn post

Figure 3.1: Experimental procedure and design.(A) Participants learned associations between visual stimuli (learn block) before and after completinga target detection task (pre-learning and post-learning blocks). Finally, memory was tested for theassociated pairs. (B) Trial structure of the target detection task. Participants indicated the presenceof a grayscaled patch (white outline for display purposes only) by button press, in every single trial.(C) Trial structure of the learning block, where to-be-paired stimuli were presented multiple times inrandom order. (D) Representational clustering logic. In the representational space - where distancesbetween the stimuli reflect their similarity - we hypothesized that learning would group associatedstimuli (mapped on two dimensions for display purposes).

3

54

3.2. ResultsMemory for the learnt pairs was assessed using two subsequent tests (Figure 3.5).We observed significant above chance level accuracy scores in the pair recognitiontask (mean = 90%, SD = 5%, chance level: 0.75, binomial test: p < 0.001) and near-perfect performance on the pair arrangement test (one single error across entiregroup), evincing ceiling encoding levels. Learning was preceded and followed byidentical target detection blocks, with identical stimulus presentation schemes (Fig-ure 3.1A-C). Overall target detection rates were high (pre-learning: mean = 98.9%, SD= 4%, post-learning: mean = 99.3 %, SD = 1%) with no significant differences betweenblocks (p = 0.99, two-tailed, Bayes Factor (BF)₁₀ = 0.26, evidence for null-hypothesisBF₀₁ = 3.9). These behavioral results suggest that participants were able to sustain asufficient and constant level of attention throughout the experiment.The target-detection blocks allowed us to inspect the change of neural stimulus rep-resentations as a function of learning, using data independent from the encodingphase. We adopted RSA to assess neural similarity between stimuli, before and af-ter learning (Kriegeskorte et al., 2006). Here, we modeled the covariance of mul-tivoxel activation patterns across stimuli, to obtain pair-wise similarity estimates(Figure 3.2, see section 3.4 for details). We subsequently visualized these similar-ity measures in a two-dimensional representational space, where the more repre-sentationally similar stimuli are located closer together. We predicted that neuralrepresentations of paired stimuli would emerge more similar compared to unpairedstimuli i.e. cluster in representational space (Figure 3.1D). We investigated two pre-defined regions-of-interest (ROIs): left and right hippocampus. For each individualstimulus, we extracted the similarity with its paired associate (within-pair similarity),which we compared to the similarity with all other, non-associated stimuli (between-pair similarity). We converted this difference score into a statistic, summarizing theamount of representational clustering in each participant (Figure 3.3A).We observed that, across our group of participants, within-pair similarity showed anincrease from pre-learning to post-learning, compared to between-pair similarity, inthe left hippocampus (Figure 3.3B, T₂₃ = 2.69, p = 0.006, BF₁₀ = 7.7). We observed asimilar, albeit non-significant pattern in the right hippocampus (T₂₃ = 0.76, p = 0.22,BF₁₀ = 0.4). Two stimuli from the target-detection blocks did not feature the learningphase, leaving them unpaired (hereafter referred to as singletons). For these single-ton stimuli we found no significant representational change, neither in positive (i.e.clustering) nor negative (i.e. segregation) direction (Figure 3.6, p > 0.17, two-tailed).To conclude our group-level analysis, we used a searchlight procedure to reveal thespatial distribution of representational clustering in spherical ROIs across the wholebrain. We observed a global peak in left amygdala (x,y,z = [-18, 0, -20] T₂₃ = 4.48, p =0.08 whole-brain FWE-corrected), extending into the hippocampus (Figure 3.3C).Next, we investigated the potential to track successful encoding of memories using

3

55

Y

betas×

voxels

betas×

compcomp×

voxels

Z U + E=

comp×

comp

G

B

A

Figure 3.2: Reconstructing representational space.(A) Pattern component model. Pattern components U were modeled using beta estimates from multi-voxel region-of-interest Y and a known design matrix Z (including a component for each stimulus,see section 3.4 for details) plus an unknown error term E. (B) Mapping representational space. Wetransformed the covariance matrix of pattern components G to a stimulus-by-stimulus dissimilaritymatrix and applied multidimensional scaling to map each stimulus in a 2-dimensional space, wherepair-wise distances reflect the dissimilarity values.

the individual reconstructed representational space. In an exploratory step, we vi-sualized the individual representational spaces of the two participants that showedthe strongest and the two that showed the weakest representational clustering inthe left hippocampus (Figure 3.4A). Although some degree of clustering in the indi-vidual spaces can be noted, manual identification of pairs among unlabeled stimuliremains challenging, even in the participants showing the strongest effect across theentire group. To quantify the utility of individual representational data, we appliedtwo classification algorithms designed to identify the four pairs from the total set.First, we employed an algorithm that incorporated prior information (i.e. the factthat there are four pairs and two singletons). The algorithm would find the config-uration of pairs yielding the highest representational clustering score (Figure 3.4B).We assessed the recall score (hit rate) of this predicted pair configuration (i.e. howmany pairs and singletons are correctly identified) for each participant (Figure 3.4C).For a majority of participants, this procedure resulted in zero correctly identifiedpairs and singletons. On average, the classification accuracy of paired and singletonstimuli did not significantly differ from chance, both in the left (p = 0.24 for pairs, p= 0.88 for singletons) and right (p = 0.75 for pairs, p = 0.23 for singletons) hippocam-pus. Secondly, we applied an algorithm that classified stimuli as being associated

3

56

L R

**

C

A B

sim

ila

rity

ch

an

ge

(z)

ne

ura

l sim

ila

rity

(z)

0.3

0.7

-1

1

-1

1

0.3

0.7

A B

pre post

**

regions-of-interest

within vs. between

.005

.05

p

Figure 3.3: Representational clustering in hippocampus.(A) For the left and right hippocampus, we compared neural similarity of associated stimuli (withinpairs, red edge) with all other stimuli (between pairs, orange edges), during the pre-learning block andthe post-learning block separately. (B) The left hippocampus showed a significant increase in neuralsimilarity within pairs, compared to between pairs, as a consequence of associative learning. ** p <0.001 for the interaction. Right: pre-learning to post-learning similarity change for both conditions,each dot represents a participant. (C) Whole-brain searchlight map. Hot colors indicate regional rep-resentational clustering. White boundaries indicate field of view. Maps thresholded at p < 0.05 forcomparison with region-of-interest results.

or not, according to a varying similarity threshold (Figure 3.4D). By applying multi-ple different thresholds, we obtained a precision-recall curve for the left and righthippocampus (Figure 3.4E), to explore signal-detection abilities and operating pointtrade-offs of the classifier (see section 3.4 for details). On average, performance didnot significantly exceed performance levels of a random classifier (95% confidenceintervals span baseline at all possible recall values).

3.3.DiscussionIn this study, we demonstrated a change in neural representations following explicitpaired-associate learning. We showed that, after learning, neural patterns of asso-ciated stimuli emerged more similar than non-associated stimuli. This representa-tional clustering effect was observed in the left hippocampus, across a group of par-

3

57

B

D

C

A

E

z = 2.51

constrained

threshold-based

z = 1.37 z = -.67 z = -.71

1

0

recall

pairs singles

1

0

recall

pairs singles

0.4

0

precision

recall 10

0.4

0

precision

recall 10

best worst

Figure 3.4: Tracking individual learning in hippocampus.(A) Representational spaces from the four participants showing the strongest and weakest represen-tational clustering (post-learning minus pre-learning), as indexed by the Z-statistics, from the left hip-pocampus. (B) Logic of constrained classification algorithm: for each participant the configuration offour pairs with the maximal within-pair versus between-pair neural similarity was chosen and servedas prediction of the learned associations (C) Pair and singleton stimulus identification accuracy (re-call) for the left and right hippocampus. Width of the colored horizontal bars reflects the proportionof participants with a particular recall score. Red line: group mean recall. Light gray shaded area withdotted line: group-level null-distribution with 95% bounds and mean. A red lines inside the shadedarea indicates that the group mean recall did not significantly differ from chance level at = 0.05. (D)Logic of threshold-based classification algorithm: stimuli are classified at a certain distance threshold.At the most strict threshold, only the two most clustered stimuli are identified as a pair, allowing forhigher precision. Conversely, at the most liberal threshold, all stimuli are marked as pairs, accomplish-ing perfect recall at the expense of precision. (E) Precision-recall curves obtained by classifying atvarious thresholds. Shaded area: 95% confidence intervals. Dotted line: null-classifier baseline. Whenthe light gray shaded area includes the dotted line, there is no significantly difference from chancelevel at = 0.05.

3

58

ticipants.Our findings complement earlier work on representational changes as a function oflearning (Visser et al., 2011; Schapiro et al., 2012, 2013, 2015; Schlichting et al., 2015;Milivojevic et al., 2015; Collin et al., 2015). Here, we investigated the effects of explicitpaired-associate learning with relatively simple visual stimuli and a well-understoodmemory paradigm. These stimuli were engineered to have minimal prior associa-tions with known concepts. We showed that the representational clustering effect,as measured with differential RSA, is replicable across memory paradigms, from themost basic to the most complex.Two earlier studies that investigated changes in neural patterns similarity reporteda representational segregation effect for non-associated stimuli (Milivojevic et al.,2015; Collin et al., 2015). However, here, we did not observe such a representationalchange for the isolated singleton stimuli. The absence of this representational seg-regation effect might be explained by two possible causes. Firstly, these previousstudies used complex narrative stimuli, wheres in the current study, we used rela-tively simple visual stimuli. Secondly, we might lack the statistical power to detect apotential representational segregation effect. Further research is required to resolvethis issue.Our observation that representational reconfigurations take place in the hippocam-pus are in line with computational models (Marr, 1971; Damasio, 1989) and experi-mental work (see Chapter 2) implicating the hippocampus a key mnemonic conver-gence zone. An array of lesion studies in animals and humans have emphasized thecrucial role of the hippocampus for associative memory (Bunsey and Elchenbaum,1996; Scoville and Milner, 1957), complemented with neuroimaging work (Eldridgeet al., 2000; Davachi et al., 2003; Rissman andWagner, 2012). In addition, the discov-ery of various specialized cell types in the hippocampus, such as place cells (O’Keefeand Dostrovsky, 1971), object-location cells (Moita et al., 2003; Komorowski et al.,2009) and concept cells (Quiroga et al., 2005), provides a neurophysiological basisfor its crucial role for associative memory. In an exploratory whole-brain analyses,we found representational clustering in a distributed set of other brain regions, inaddition to the hippocampus. This observation accords with previous reports of rep-resentational clustering in regions such as the insular cortex (Schapiro et al., 2013)and mPFC (Milivojevic et al., 2015).But how might the representational clustering effect be realized on a cellular level?Electrophysiological recording studies have shown that stimulus-specific firing pat-terns emerge as a function of learning (Sakai and Miyashita, 1991; Ison et al., 2015).Following paired-associate learning, neurons in the medial temporal lobe have beenshown to selectively encode associations, even after only few learning exposures.But how might this cellular effect be measurable on a population level, using acoarse neuroimaging technique with a limited spatial resolution, such as fMRI? Evi-

3

59

dence from intracranial recordings in animals suggest that neighbouring neurons inthe medial temporal lobe develop similar response preferences, as a result of learn-ing (Erickson et al., 2000). This clustering of neurons with similar firing patternspotentially allows for a population bias, which can be picked up with non-invasiveneuroimaging techniques and multivariate pattern analyses.In addition to the described group-level effects in the hippocampus, we investi-gated representational clustering on an individual level. First, we reconstructed themnemonic network structure of individual participants, using differential RSA. Next,we assessed the utility of these reconstructed networks for the purpose of track-ing learning. Here, we applied two different classification approaches. Firstly, weemployed a custom algorithm with prior information on the amount of pairs andsingletons, designed to find the most optimal configuration given the neural simi-larity data. Secondly, we used a threshold-based approach to map the sensitivityand precision of information present in the neural similarity data. With the currentmethodology however, we were unable to identify the learned associations formneural similarity data. Neither the prior information nor a signal-sensitive threshold-based algorithm yielded significant predictive value. The fact that we did find a sig-nificant hippocampal representational clustering effect on the group level indicatesthat significant variables are not necessarily predictive (Lo et al., 2015). Neverthe-less, our dataset and methods can serve as a benchmark to aid the developmentof improved individual associative memory network reconstruction methods. Withimproved reconstruction methods, recovery of the learnt information from neuraldata may still be possible. Importantly, the approach can potentially be leveragedto read-out a the memory or knowledge network of an individual and subsequentlyapplied in educational settings to track the acquisition of new knowledge.In sum, we have shown that hippocampal neural pattern similarity increases fornewly associated stimuli. We demonstrated this representational clustering effectdue to learning in the context of a well-understood associative learning paradigmwith basic stimuli. Although more research needs to be done to successfully recon-struct individual associative memory networks, our work can be viewed as a valida-tion of the differential RSA method for further investigations into the neural codingprinciples underlying memory.

3.4.Methods3.4.1. ParticipantsA total of twenty-six healthy volunteers participated in the study (16 female, aged 18-28 years,average: 23 years). We performed a screening to ensure only individuals with no history ofneurological disease, normal or corrected to normal vision and no color blindness were in-cluded. All participants gave informed consent and were reimbursed for partaking in theexperiment. The local ethical review committee (CMO committee on Research Involving

3

60

Humans, region Arnhem-Nijmegen, the Netherlands) reviewed and approved all experimen-tal procedures. After data acquisition, two participants (both female) were excluded: onedue to excessive head motion (more than 4 mm absolute head displacement) and anotherdue to incomplete data (technical difficulties).

3.4.2.MaterialsWe used custom-made, symmetrical, colorful, fractal-like visual stimuli, created with theGimp software package (v2.6, www.gimp.org). The abstract nature of these stimuli was de-liberately chosen tominimize the influence of any prior associations participants might havewith the to-be-associated material. We created a collection 40 stimuli in total, from whichwe selected a subset of the ten most dissimilar stimuli, using the results of a behavioral ex-periment performed in an independent group of five pilot participants. Here, we assessedthe across-stimuli similarity structure, using the ARENA procedure (Kriegeskorte and Mur,2012), where participants were instructed to repeatedly arrange the 40 stimuli in a circleon screen, according to their judged similarity. We intentionally provided no instructionto what feature (e.g. color, associations, pattern) the similarity was to be judged, in orderto obtain an unbiased estimate of the average inter-stimulus relations. The ARENA proce-dure yielded an average dissimilarity matrix of the 40 stimuli, from which we selected theten most dissimilar ones: we performed a jackknife procedure and searched for subset often stimuli with a maximal Sharpe ratio (dissimilarity mean-variance ratio). This procedureyielded dissimilar stimuli, but with low variance across their similarity scores (i.e. equidistantin similarity space and not in a single cluster). Finally, we manually checked whether thissubset was dispersed across the entire set instead of within one grouped stimuli. The sameten stimuli were used for all participants in the main experiment.

3.4.3. Experimental taskWeused a basic paired-associate learning task (Figure 3.1C), wherewe instructed participantsto remember a set of paired stimuli. Using eight out of the ten stimuli, we constructed fourpairs at random for each participant, leaving two stimuli as unpaired controls (singletons).Each pair was presented eight times, organized in blocks. In each trial, the paired stimuliwere shown next to each other for 8000 ms, followed by a 2000 ms inter-trial interval (ITI).In each block, all pairs were presented once, in random order. In total, the learning phasespanned approximately 6 minutes. The learning phase was preceded and followed by twoidentical target detection task blocks (Figure 3.1B) (Schapiro et al., 2012, 2013). Here, stim-uli were repeatedly shown for 1000 ms, against a gray background, followed by a variablelength ITI (short: 1000ms, medium: 3000ms, long: 5000ms). In 10% of all trials, a small por-tion of the stimulus (about one sixth of the pixels) was rendered in grayscale. Participantswere instructed to detect these target events and respond by pressing a designated button.In case there a given trial was not a target event, participants pressed the alternative non-target button, requiring them to respond in every trial and remain attentive. The mappingof the target and non-target buttons was swapped across participants to avoid response bi-ases. Each stimulus was presented 28 times in random order inminiblocks. The presentationscheme of these miniblocks was carefully counterbalanced and optimized for modeling anddifferential RSA: the varying ITI lengths (short, medium, long) and targets were distributedequally across stimuli. Next, we searched for a suitable candidate presentation scheme with

3

61

the following constraints: in each miniblock, each stimulus should be presented once andthe duration of all miniblocks should not differ too much. The maximum standard devia-tion of a single miniblock duration, in related to the mean duration across miniblocks, wasset to 1.0. There was no constraint on the number of target trials within each miniblock, tokeep the occurrence of target trials unpredictable. The onset of the stimuli was time-lockedto the scanner sequence. One entire target detection block of approximately 20 minuteswas divided in six subsections, with performance feedback and a pause in between. Impor-tantly, the pre-learning and post-learning target detection blocks had identical presentationschemes, including ITI lengths and the target patch location. The performance data for thetarget detection task of 21 out of 24 participants was used, due to technical difficulties with 3participants. The post-learning target detection blockwas followed by a forced-choice recog-nition test outside the scanner. Here, a mix of paired and unpaired stimuli was presentedin a one-to-three ratio. Each specific combination of stimuli was repeated eight times. Par-ticipants were instructed to indicate as fast as possible and as accuracy as possible whetherthe two presented stimuli were paired in the learning phase. Again, response buttons wereassigned randomly across participants. Finally, we administered a pair arrangement test todouble-check successful learning. Here, participants were shown all ten stimuli on screen,randomly arranged in a circle. We then asked them to drag and drop stimulus pairs into anarbitrary quadrant of choice. The two singleton stimuli should remain in gray area in themiddle of the screen, outside of the white quadrant boxes. The target detection task, learn-ing phase and recognition test were programmed with Neurobehavioural Systems Presenta-tion (v16.0 neurobs.com). Text files containing the counterbalanced stimulus presentationschemes and the ARENA task were created with MATLAB (v2014a, The Mathworks).

3.4.4. fMRI data acquisition and preprocessingWe acquired neuroimaging data using a 3T Siemens Prisma scanner in combination with a32-channel head coil. We used 2D echo-planar imaging (EPI) sequence (voxel size: 2.0 ×2.0 × 1.8 mm, volume TR: 2000 ms, TE: 24.0 ms, flip angle: 85 degrees, 37 slices, distancefactor: 11%, field of view (FOV): 210 × 201 × 74 mm, orientation: -17.8 degrees from thetransverse plane), designed to obtain relatively fast and high-resolution data of the medialtemporal lobe and prefrontal areas. The Siemens ”AutoAlign” head software was used tokeep the FOV tilt constant across participants. To aid registration procedures, we acquiredan additional whole-brain snapshot at the end of the experiment (75 slices, FOV: 210 × 210× 150 mm, TR: 5290 ms, 5 volumes, other acquisition parameters identical to 2D EPI se-quence above). Also, prior to the experiment, we acquired a T1-weighted structural using astandard MPRAGE-grappa sequence (voxel size: 1.0 mm isotropic, volume TR: 2300 ms, TE:3.03ms, flip angle: 8 degrees, 192 slices, distance factor: 50%, FOV: 256 × 256 × 192 mm).Preprocessing of the fMRI data was performed with tools from FSL v5.05. Functional imageswere brain-extracted (Smith, 2002), motion-corrected to the middle volume using MCFLIRT(Jenkinson et al., 2002) and high-pass filtered (50 s cutoff, Gaussian sigma: 12.5). We alsobrain-extracted the middle volume of the whole-brain functional for later use as registra-tion reference. The structural images were brain-extracted and segmented into gray matter,white matter, cerebrospinal fluid (CSF) and out-of-brain voxels, using FAST (Zhang et al.,2001). With these structural masks, we extracted the mean compartment signals from thefunctional sessions for later modeling purposes. In addition to the FAST segmentation, we

3

62

performed a more fine-grained automated segmentation of the subfields of the hippocam-pus (Leemput et al., 2009) using FreeSurfer v5.3. We used the preprocessed structural scansof the entire group of participants to create a group-specific template with the ANTs toolbox(Avants et al., 2011) v2.10. The following registration parameters were estimated: functionalsessions (pre-learning, learning and post-learning) to reference functional (FLIRT, 6 degreesof freedom (DOF), fine search), reference functional to structural (FLIRT, 6 DOF), structuralto group template (ANTs build template script with default parameters), and group templateto 1 mm Montreal Neurological Institute standard brain (ANTs: rigid, affine and deformablesymmetric normalization).

3.4.5. Representational similarity analysisWe used a modeled data as input for our RSA. Therefore, we fitted a General Linear Model(GLM) to the time series data, using FILM generalized least squares with prewhitening(threshold: 1.0, smoothed autocorrelation estimates, susan mask size: 5). Conditions-of-interest were modeled in the following way: we grouped the 25 trials per stimulus basedon their ITI length. Each grouping contained at least one short, one medium and one longITI trial. The remaining trials were distributed equally among the initial triplet groupings,resulting in six to eight groupings per unique stimulus (variations across stimuli due to therandom component of the counterbalancing scheme described above). The events within agrouping were distributed uniformly across time, to counter non-specific temporal effects,such as habituation and noise autocorrelation. The onsets and durations of grouped eventswere entered together as one regressor in the design matrix for the GLM. In addition, weincluded the six head movement parameters estimated during motion correction and threemean compartment signals (white matter, CSF and out-of-brain) as nuisance regressors. Thefinal design matrix was high-pass filtered and convolved with the default double-gammaHemodynamic Response Function. We modeled the data in the individual functional ses-sion space, in order to reduce registration deformations to a minimum. We did not smooththe functional data prior to modeling, in order to retain potential fine-grained across-voxelpatterns. The resulting beta coefficients of each regressor were registered to the referencefunctional image, before entering the RSA. All subsequent analyses were programmed withMATLAB.For the region-of-interest RSA, we created participant-specific gray matter masks of theleft and right hippocampus by combining the hippocampal subfields segmentations fromFreeSurfer. Voxel patterns were subsequently extracted for each of these ROIs. To estimatethe representational similarity of our stimulus conditions, we performed pattern componentmodeling (Figure 3.2) (Diedrichsen et al., 2011). In thismultivariatemodeling framework, mea-sured patterns are decomposed into their constituent parts, allowing one to more reliablyestimate their true correlations. We modeled the following components: a generic com-ponent (constant) including all regressors, two session components (pre-learning and post-learning) including all regressors of a single session and twenty stimulus components fromboth the pre-learning and post-learning session. Each stimulus component encompassedits own six to eight beta estimates. We used the following constraints when estimating thecovariance matrix of the pattern components: the generic component was set to be orthog-onal to all other components (e.g. uncorrelated to a specific stimulus pattern component),the session components were set to be orthogonal to all stimulus components (e.g. session

3

63

components may be correlated, but should be uncorrelated to stimulus components) andvariance of the stimulus components was fixed (e.g. stimulus components should all havesimilar variance magnitude). We fitted the pattern component model with default hyper-parameters. Subsequently, we converted the resulting 23-by-23 variance-covariance matrixto a correlation matrix, by dividing the covariance between two components by their vari-ance product. We Fisher Z-transformed correlation matrix and subtracted the pre-learningfrom the post-learning similarity values. Finally, we computed the average within-pair mi-nus the average between-pair similarity (Figure 3.2A). We used a permutation-based null-distribution to convert the observed difference score from the within-between comparisonto a Z-statistic, summarizing the amount of representational clustering. To obtain this null-distribution, we recomputed the within-between difference score for each possible permu-tation of the correlation matrix. Next we calculated the probability of observed statisticgiven the null-distribution from permuted data. This p-value was converted to a Z-statistic,yielding a single value per ROI, for each participant. Next, we tested if the Z-scores weresignificantly above zero, using a one-tailed non-parametric paired T-test with 10,000 per-mutations. In addition, we calculated Bayes Factors using the standard implementation ofthe Bayesian Paired Samples T-test in the JASP software package (v0.7.1.12, jasp-stats.org) toobtain an indication of howmuchmore likely our hypothesis (i.e. higher similarity for pairedstimuli than unpaired stimuli, from pre-learning to post-learning) is than the null hypothesis(i.e. no difference). In addition to the ROI analysis, we performed an exploratory whole-brainsearchlight analysis. Here, the multivoxel activity pattern from each spherical searchlight (3voxel radius, minimum of 30 gray matter voxels), was extracted from the modeled beta im-ages, for each condition. A pattern componentmodel, identical to ROI analysis, was fitted foreach sphere and the resulting similarity scores were mapped back to the center voxel. Sub-sequently, we Fisher Z-transformed these correlation maps and subtracted the pre-learningfrom post-learning similarity values. The resulting similarity difference maps were furtherreduced to a single image containing the average within-pair minus the average between-pair similarity for each voxel. We warped the difference maps to 2 mmMNI space, using thepreviously estimated registration parameters. After smoothing the maps (FWHM: 6 mm)we used RANDOMISE (Winkler et al., 2014) to test which areas in the brain showed highersimilarity for paired stimuli than unpaired stimuli (non-parametric, 10,000 permutations)on the group level. The test statistic was based on a one-sided T-test, with 5 mm variancesmoothing and threshold-free cluster enhancement (TFCE) (Smith and Nichols, 2009). Theresulting p-values were thresholded at p < 0.05 uncorrected, log-transformed and renderedon whole-brain cortical surfaces using the CARET brain visualization software package, v5.65.

3.4.6. Individual representational space analysisFor the individual representational space analysis, we took the individual Fisher Z-transformed correlation matrices from the hippocampal ROI pattern component model re-sults. Next, we converted these correlation matrices to dissimilarity matrices, by takingthe square root of the one minus similarity scores. To visualise the stimulus dissimilaritiesin two dimensions, we employed metric multidimensional scaling (stress criterion, randomstarting position). The predictive value of the Fisher Z-transformed correlation matrices wasassessed using two algorithms: first, we constructed a constrained classification algorithm,that incorporated prior knowledge on the pairing structure. This algorithmwas programmed

3

64

to derive the configuration of four pairs with the maximal within-pair versus between-pairneural similarity for each participant, by going through all possible permutations of the cor-relation matrix. This configuration was chosen as the final prediction of learned associa-tions and allowed us to compute a recall score (hit rate) per participant (e.g. pair recallof 1.0 implies that four out of four pairs were correctly identified, likewise for singletons).We then averaged recall scores across participants and assessed the probability (p-value) ofthis observed average recall score using a permutation test. Here, we randomly shuffledthe pair identities (10,000 permutations) to obtain a null-distribution indicating how manypairs and singletons are identified correctly by chance (guessing). The second algorithm didnot incorporate prior knowledge on the number of paired stimuli, but instead tested thepositive predictive value (precision: the fraction of true positives among positive calls) ofthe neural similarity data as a function of sensitivity (recall: true positive rate). First, werank-transformed the correlation matrices (e.g. the highest similarity value gets assignedthe highest rank) and concatenated the resulting rank scores of all participants. Next, thealgorithm classified the stimuli as being associated (positive call) at various rank thresholds.To illustrate this principle, we consider the most liberal and most strict threshold: at themost strict threshold, only the two most similar stimuli are identified as a pair. In case thesetwo stimuli were actually paired, we obtain a high precision score. Conversely, at the mostliberal threshold, all stimuli are marked as being associated. In the latter case, we obtainperfect recall (all pairs are identified correctly), but we necessarily have a very low precisionscore. In the ideal case, the within-pair similarity values are higher than all other values inthe correlation matrix. This signal-sensitive procedure allowed us to create precision-recallcurves for each participant. The 95% confidence intervals of this curve were obtained usingbootstrap resampling (1,000 iterations) and compared to a null-classifier baseline (i.e. themean recall score when guessing).

3

65

3.5. Supplemental Information3.5.1. Supplemental Figures

A B

Figure 3.5: Memory tests.(A) Experimental design of the forced-choice recognition test. Participants were shown combinationsof stimuli and indicated by button press whether the stimuli were paired in the learning phase. (B)Starting screen of the free pair arrangement test, with the stimuli arranged randomly in circle. Partic-ipants were instructed to drag and drop the paired stimuli in a single quadrant. Unpaired singletonscould be left in the center of the screen.

-1

2

pre post -1

2

pre post

ne

ura

l sim

ila

rity

(z)

Figure 3.6: Representational changes of singleton stimuli.For the stimuli that were not shown during the learning phase and therefore not paired with anyspecific other stimulus, we did not observe any representational effect of the learning phase. Bothin the left and right hippocampus, the similarity between singletons did not significantly increase ordecrease.

Mesologue

November 9th, 2013, San Diego, California, USA.

J&B straight, and a Corona. My buddy Sander ordered a drink at theHilton hotel bar. Our group leader, Christian, decided to order aWeizen,after going through the specials menu. Then it happened: the famouselectrophysiologist walked by. Hewaved atme. Prompted by this event,Sander remembered that the electrophysiologist gave a keynote lectureat the satellite workshop, as he had noted from an earlier glance atthe conference programme. In addition, Sander was reminded of thefact that I had attended the Meet The Speakers dinner cruise two daysago. At that moment, using these two separate pieces of information,Sander inferred that the electrophysiologist and I must have gotten ac-quainted during the dinner. From this moment onwards, his memoriesof me were, to a limited albeit significant extent, associated with theelectrophysiologist, linked together by the Hilton event. He effectivelyrecalibrated and reconfigured his memories. Sander wondered: howwas he able to perform this remarkable feat?

67

4Hippocampal-Prefrontal Theta

Oscillations Support Memory Integration

Alexander R. Backus, Jan-Mathijs Schoffelen, Szabolcs Szebényi,SimonHanslmayr, Christian F. Doeller

This chapter is published as: Backus, A.R., Schoffelen, J-M., Szebenyi, S., Hanslmayr, S., Doeller, C.F.(2016). Hippocampal-prefrontal theta oscillations support memory integration, Current Biology, 26:1–8, doi: 10.1016/j.cub.2015.12.048

69

4

70

4.1. IntroductionDuring everyday life, we continuously bind new information into coherent episodicmemories (Eichenbaum et al., 1999). Although these memories are inherently sep-arated in time, we have the remarkable ability to link and recombine episodes withoverlapping elements (Preston and Eichenbaum, 2013; Zeithamova et al., 2012a; Ku-maran et al., 2009). Integration of multiple events into a new memory forms thebasis of inferential reasoning (Eichenbaum et al., 1999), regularity learning (Doelleret al., 2005), and decision making and ultimately the formation of our knowledgebase (Kumaran et al., 2009).Evidence from animal lesion studies (Bunsey and Elchenbaum, 1996) and humanneuroimaging (Zeithamova et al., 2012a; Horner et al., 2015; Shohamy and Wagner,2008; Zeithamova and Preston, 2010; Schlichting et al., 2015; Milivojevic et al., 2015;Collin et al., 2015) has demonstrated that the medial prefrontal cortex (mPFC) andhippocampus (Preston and Eichenbaum, 2013) are the two key regions implicatedin memory integration. Interestingly, human functional MRI (fMRI) studies haverevealed increased functional connectivity between these two key nodes duringmemory encoding and retrieval, including integrating information across events (Zei-thamova et al., 2012a). However, due to the low temporal resolution of fMRI, theelectrophysiological mechanisms underlying this crosstalk by which the hippocam-pus andmPFC are able to retrieve, exchange, integrate, and re-encodemultiplemem-ories on a millisecond timescale remain poorly understood.Rhythmic theta band activity in the hippocampus (traditionally 4–8 Hz in humans,6–10 Hz in rodents), which is strongly associated with place cell activity (O’Keefeand Recce, 1993), has been implicated in memory formation by intracranial record-ing studies (Lega et al., 2012), although human studies commonly report effectsat the lower end of the traditional theta band (Jacobs, 2014; Watrous et al., 2013).More recently, these findings have been corroborated by studies using magnetoen-cephalography (MEG) (Cornwell et al., 2008; Guitart-Masip et al., 2013; Kaplan et al.,2014; Staudigl and Hanslmayr, 2013), supported by modeling and invasive recordingefforts that confirm the feasibility of reconstructing hippocampal theta oscillationsfrom MEG sensor data (Dalal et al., 2013).In addition, interregional coupling of theta oscillations in the hippocampus andmPFC has been observed during memory encoding, retrieval, and decision mak-ing in animals (Brincat and Miller, 2015; Siapas et al., 2005) and humans, using in-tracranial recordings (Anderson et al., 2010) and MEG (Kaplan et al., 2014). Suchoscillatory coupling between distant regions has been put forward as an electro-physiological mechanism for information transfer (Fell and Axmacher, 2011). Takentogether, these findings suggest that theta oscillations might be involved in orches-trating the integration of memories. Theoretical models and recent neuroimagingevidence suggest that memory integration is achieved through retrieval-mediated

4

71

learning (Zeithamova et al., 2012a; Shohamy and Wagner, 2008). Since theta oscil-lations have been posited to gate information flow during alternating encoding andretrieval states (Hasselmo et al., 2002), we hypothesize that rhythmic theta band ac-tivity plays an important role duringmemory integration, where an existingmemoryis retrieved and re-encoded together with a new memory.In sum, while the mPFC and hippocampus appear to play a crucial role in integrat-ing multiple memories, the underlying electrophysiological mechanism remains un-clear. Synchronized theta oscillations are likely to provide such a mechanism, buttheir region-specific involvement in human memory integration remains elusive. Toresolve this outstanding issue, we used MEG to record whole-brain oscillatory ac-tivity of participants performing a classic associative inference paradigm (ShohamyandWagner, 2008). We leveraged novel source reconstruction methods to measurehippocampal theta oscillations and employed coherence analysis to investigate os-cillatory coupling with the mPFC. Critically, we aimed to pinpoint electrophysiolog-ical markers during encoding of novel information that are predictive of successfulintegration with an existing memory.

4.2. ResultsParticipants performed an associative inference task (Zeithamova and Preston, 2010)modified for MEG, in which pairs of to-be-associated object stimuli were briefly pre-sented in sequence (see section 4.4 for details). Pairs comprised so-called premiseassociations (AB and CB pairs) and a control association (YX pair). Crucially, par-ticipants were asked to subsequently infer an indirect, unseen link (AC association)between the overlapping AB and CB pairs (Figure 4.1) and thus encode a collectionof triad (ABC) and dyad (YX) memories. Following encoding, we tested the partici-pant’s memory for all associations. On average, participants correctly remembered79.8% (SEM = 2.8%) of AB pairs, 75.0% (SEM = 3.7%) of YX pairs, 69.0% (SEM = 3.8%) ofCB pairs, and 62.3% (SEM = 3.9%) of the crucial inferred AC associations (Figure 4.2A).We observed a clear pattern across different association types: the second premisepairs (CB) were remem- bered significantly worse than the initial AB premise pairs(T₂₆ = 8.13, p = 1-⁵ Bonferroni-corrected [corr]) and control YX pairs (T₂₆ = 3.81, p =0.006, corr). In turn, performance on directly associated objects, including the CBpairs, significantly surpassed inferred AC associations (T₂₆ = 4.75, p = 0.0004, corr).Next, we excluded seven participants from subsequent MEG analyses, who were un-able to reach the performance criterion on AC association tests (see ExperimentalProcedures for details). Based on final performance, we categorized each individ-ual triad into eight possible categories, ranging from ”no links remembered” to ”alllinks remembered” (Figure 4.2B). Through behavioral piloting, we had adjusted taskdifficulty to obtain roughly half of the triads in the ”all links remembered” category(mean = 56.5%, SEM = 3.8%).

4

72

encoding

block 1

memory

test

encoding

block 2

run 12run 1 run 2 run 3

2-2.5s

0.2s

0.2s

4s

S1

0.2s

S2encoding

interval 1.5-2s

0.3s

1.5s

max 5s

max 3s

cue

probes

con!dence?

1 2 3 4

Figure 4.1: Experimental Procedure and Trial Structure of theMemory Integration Task.Top: in 12 cycles, participants learned dyad (YX) and triad (ABC) associations between gray-scale pic-tures of objects during two separate encoding blocks. Subsequently, memory was probed for bothdirectly associated objects (AB, CB, YX) and inferred associations (AC). Bottom: each encoding trialcomprised serial presentation of two objects (S1 and S2: first and second stimulus), followed by a ded-icated encoding interval. A red fixation cross indicated a short blink phase and the upcoming newtrial. Test trials commenced with a cue, a retrieval phase, a forced-choice response phase with fouralternatives and concluded with a memory confidence rating.

To test our primary hypothesis that hippocampal theta oscillations are involved inmemory integration, we applied a ”subsequent integration contrast” (Figure 4.3A).Here, we compared brain activity during CB encoding trials where the AC associ-ation was later successfully integrated, with a subset of encoding trials where theCB premise or XY association was remembered, but, crucially, no indirect AC linkwas inferred. By including brain activity related to direct associative encoding ofthe premise pair in the non-integration subset, we aimed to isolate processes con-tributing tomemory integration. After removing effects due to eye-movements (seesubsection 4.5.2 for control analysis) and other artifacts from the signal, we pursueda novel, advanced region of interest (ROI) source reconstructionmethod to estimatetheta power from the left and right hippocampus. In particular, we applied leadfieldreduction based on anatomical priors (see section 4.4 for details and Figure 4.5 fora graphical depiction) where we took into account the structure of the hippocam-pus. Initially, we targeted a broad frequency range of theta oscillations spanning 3-7Hz—a slightly lower frequency than the traditional theta band, based on recent re-ports (Jacobs, 2014; Watrous et al., 2013). Using a sliding time window, we obtained

4

73

0

1

N = 20

0

100* **

N = 27

****

****

***n.s.

pe

rfo

rma

nce

(% c

orr

ect

) p

rop

ort

ion

A

B

Figure 4.2: Behavioral Performance.(A) Across different association types, performance for premise pairs was better than for inferred pairs.Schematics below bars depict different conditions (e.g., AB nodes with an edge symbolize AB paircorrect). Red line: mean, darker shaded area: SEM, dotted line: accuracy chance level, dashed line:exclusion criterion, red-circled dots: excluded participants, *p < 0.01, **p < 0.001, ***p < 0.0001, ****p <0.00001. (B) Proportion of triad associations in each fine-grained performance category (see schematicbelow bars). Each dot represents data from a single participant (A and B).

the time course of theta power and converted the values to normalized differencescores (T-statistics) for the subsequent integration contrast, separately for the leftand right hippocampus. Since previous electrophysiological work has demonstratedthat memory retrieval and encoding occurs rapidly (Ranganath and Paller, 1999), wefocused our initial statistical test on the first two seconds of the encoding interval.We found a significant difference in theta amplitude (p = 0.04 cluster-corrected),where power in the left hippocampus was increased from 350 to 1,000 ms follow-ing stimulus offset in successful integration trials, compared to non-integration tri-als (Figure 4.3B). Overall, the right hippocampus showed a similar pattern of thetapower differences over time, albeit non-significant (power increase: p > 0.31 cluster-corrected). The left hippocampal theta increase peaked at 400ms into the encodinginterval (Figure 4.3C, T₁₉ = 2.58, p = 0.007, Bayes Factor (BF₁₀) = 6.1, see section 4.4for details). Note that due to the applied estimation procedure, this effect contains

4

74

data from a 1-s time window, spanning -100 to 900 ms.In a next step, we performed a frequency-resolved follow-up analysis to display thespectrotemporal specificity of the described early theta difference (Figure 4.3D; seeFigure 4.6 for the right hippocampus). In addition, we corroborated results from theleft hippocampus with an alternative source reconstruction algorithm (Figure 4.7)and sensor level data showing a similar pattern in temporal sensors (Figure 4.8; seeFigure 4.9 for an exploratory analysis of other frequency bands). Finally, we esti-mated theta power of a whole-brain source grid at the peak time window and com-puted difference scores with the subsequent integration contrast (Figure 4.3E). Asexpected, we observed a significant difference between conditions (p = 0.01, whole-brain cluster-corrected) with a spatially specific cluster in the left hemisphere (peakof cluster in middle temporal gyrus, Brodmann area 21; x, y, z = [-76, -24, -16], T₁₉= 3.92, extending into the hippocampus). In addition, we observed a cluster in theright hippocampus (p = 0.03 whole-brain cluster-corrected; peak of cluster in supe-rior temporal gyrus, Brodmann area 22; x, y, z = [44, -16, -8] T₁₉ = 4.07, including theright hippocampus). We observed no other significant theta power increases in thebrain (p > 0.44 whole-brain cluster-corrected, see Table 4.1 for list of brain regionsthresholded at p < 0.01 uncorrected).

4

75

-3

T

3

S1-S2

C-B C-B Y-X

integrated vs. not integrated

C-B

block 2 premise

association

pe

ak

the

ta p

ow

er

(a.u

.)

0 22

12

fre

qu

en

cy (

Hz)

time (s)

*

A

B

D

C

*-3

3

2

12

E

0

T

5

the

ta p

ow

er

di!

ere

nce

(T-v

alu

e)

L R

C-B

encoding

block 2

aggregate

test result

4

76

Figure 4.3: Hippocampal Theta Power Predicts Successful Memory Integration.(A) Subsequent integration contrast: each triad or dyad was categorized according to its aggregatememory test result (top row). This categorization was used to assign the corresponding encoding trialfrom the second block to the integration (red) or non-integration (orange) condition. The middle rowshows a stream of five encoding trials, colored according to their condition assignment. Brain activityduring integration trials was contrasted with non-integration trials, controlling for direct encoding ofthe premise pair (dark brown link). (B) Normalized theta power (3-7 Hz) difference scores (T-statistics)over time for the left (light purple) and right hippocampus (dark purple) for the subsequent integra-tion contrast. Time-axis from (D), where t = 0 marks the start of the encoding interval. Horizontal barindicates significant theta power increase in the left hippocampus from 300 to 1,000 ms into the en-coding interval. *p < 0.05 cluster-corrected. (C) Peak statistics for each separate condition, where eachdot represents one participant. Colored line, mean; lighter shaded area, SEM. (D) Full time frequencyrepresentation of the left hippocampus. Red indicates integration, while blue denotes stronger thetapower during non-integration trials. White dotted lines show the statistical window-of-interest usedin (A). In order to display all data, we applied no threshold to the T-values. (E) Whole-brain spatialdistribution of theta power 400 ms into the encoding interval. Slices (x, y, z = [-33, -22, -16]) wereselected in order to visualize effects in both the left and right hippocampus. Maps thresholded atcluster-threshold value p < 0.01 for display purposes.

In a second analysis, we investigated functional coupling between the left hippocam-pus andmPFC at the peak timewindow of the theta power subsequentmemory inte-gration effect. To this end, we performed a seed-based functional connectivity analy-sis, in whichwe computed coherence across trials between the left hippocampal ROIsignal and the whole-brain grid sources (Figure 4.4A and B). We then searched forcoupling effects inside an anatomically defined area comprising the mPFC (Schlicht-ing et al., 2015). We observed a significant difference in coupling (p = 0.03 search-volume cluster-corrected), with a spatially selective cluster in the mPFC where thetaoscillations were more strongly coupled with left hippocampal theta when integra-tion was successful, compared to non-integration trials (Figure 4.4B and C, peak: x,y, z, = [-4, 40, -8]). The cluster mainly covers the left mPFC and included parts ofBrodmann areas 10, 11, and 25, with a local peak coherence in the orbital part of theleft middle frontal gyrus (T₁₉ = 2.97, p = 0.004, BF₁₀ = 12). Markedly, we found that thepeak coherence voxel did not show a significant increase in theta power (T₁₉ = 0.92,p = 0.81), with evidence suggesting that theta power levels did not differ across con-ditions (BF₁₀ = 0.13, support for null-hypothesis: BF₀₁ = 7.5). Therefore, the observedcoherence increase is unlikely to constitute a side effect of a potential overall signalamplitude increase. In addition, we observed a similar pattern of results when weused phase-locking values, a coupling measure that is less sensitive to co-variationin power between regions (Figure 4.10A). In both conditions, phase delays betweenthe left hippocampus and the mPFC peak voxel did not cluster around zero (Fig-ure 4.10B), suggesting that the observed phase coupling effects are not due to spatialleakage of activity (see subsection 4.5.2 for details). There were no other significanttheta coherence increases in the brain (p > 0.08 whole-brain cluster-corrected, see

4

77

Table 4.2 for list of brain regions thresholded at p < 0.01 uncorrected).

seed

region

A C **

0

T

3

anatomical

mask

B

5 ± 2 Hz 0.4 ± 0.5 s

0

0.6

pe

ak

coh

ere

nce

L R

Figure 4.4: Hippocampal-Prefrontal Coupling Signals Successful Memory Integration.(A) Seed-based theta coupling analysis logic. Time-frequency window of the peak theta power effectwas used to compute theta coherence of left hippocampal seed region to the rest of the brain, focusingon the anatomically delineated themPFC (see schematic ofmask). (B) Brain regions showing increasedcoherence with the left hippocampus in the subsequent integration contrast. Slices centered on thecoherence peak in the mPFC. Maps were thresholded at cluster-threshold p < 0.01 for display purposes.(C) Peak statistics for both conditions separately, where each dot represents the peak coherence ofthe left hippocampus to the mPFC of one participant. Note that although raw coherence metrics aredisplayed here, debiased Z-transformed measures were used for the significance test. **p < 0.005.

4.3.DiscussionIn this study, we have demonstrated the involvement of hippocampal and prefrontaltheta oscillations in memory integration in humans. By leveraging the high tempo-ral resolution of MEG, we showed that theta signals in the medial temporal lobeincrease in amplitude when a new memory is successfully incorporated into an ex-isting mnemonic representation.Rhythmic activity in the theta frequency band is the most prominent type of activ-ity signaling the online state of the hippocampus and surrounding medial temporalbrain regions (Buzsáki, 2002). Individual cell firing is phase-locked to theta waves,generating phase-coding and neuronal population sequences (O’Keefe and Recce,

4

78

1993). Moreover, the alternating phases of theta are implicated in rapid switchingbetween inputs and outputs of the hippocampus (Hasselmo et al., 2002; Jezek et al.,2011). This input-output gating has been put forward as a mechanism to segregateencoding and retrieval states and prevent potential interference (Hasselmo et al.,2002). A large body of evidence directly links theta to memory function: on thecellular level, rhythmic excitability modulation by theta is essential for long-termsynaptic potentiation (Capocchi et al., 1992). On the population level, theta am-plitude tends to markedly increase when novel information is encoded and storedinformation retrieved from memory, for instance, during spatial navigation (Kaplanet al., 2014) and decision making (Guitart-Masip et al., 2013). Moreover, global dif-ferences in theta oscillations both during and preceding encoding have been linkedto memory performance (Sederberg et al., 2003; Long et al., 2014; Addante et al.,2011). Interestingly, some studies report increases (Sederberg et al., 2003; Addanteet al., 2011) while others report decreases (Long et al., 2014) in theta power duringsuccessful memory encoding, leaving the precise contribution of theta to memoryunresolved. The behavioral benefits or detrimental effects of enhanced theta oscilla-tions during encodingmight highly depend on differences in encoding strategies andmemory tests across subsequentmemory studies (Hanslmayr and Staudigl, 2014). Inour data, we also observed a theta decrease in the later phase of the encoding win-dow, which could be potentially due to conflict processing in the non-integrationcondition (Oehrn et al., 2015) or enhanced information processing in the integrationcondition via oscillatory desynchronization (Hanslmayr and Staudigl, 2014). How-ever, with our hypothesis-based approach, we investigate a very specific role forincreases in theta oscillations during the integration of prior memories with newinformation, going beyond traditional subsequent memory studies.Previous electrophysiological work has demonstrated that a retrieval cue can lead toreactivation of a memory very rapidly, within 500 ms (Ranganath and Paller, 1999).In line with these reports, we showed a similar time course during memory inte-gration. The significant increase in theta oscillations 350 ms after stimulus presen-tation suggests that encoding of the inferred association (AC) immediately followsthe reactivation of the premise association (AB). This observation accords with theretrieval-mediated learning hypothesis (Zeithamova et al., 2012a). Taken together,our findings support the notion of theta oscillations as the key operating mecha-nism of the hippocampus for information processing. In particular, during retrieval-mediated learning of an integrated memory, hippocampal theta oscillations mightsubserve segregation of the necessary retrieval and encoding processes (Hasselmoet al., 2002).In addition to a hippocampal theta amplitude increase, we showed that enhancedtheta coupling between the hippocampus and mPFC predicts successful memoryintegration. Our findings are consistent with previous observations of hippocampal-

4

79

prefrontal interactions during spatial navigation (Kaplan et al., 2014), decision mak-ing (Guitart-Masip et al., 2013), and other memory tasks (Simons and Spiers, 2003).In particular, we corroborate previous fMRI studies showing the importance ofhippocampal-prefrontal crosstalk for memory integration (Zeithamova et al., 2012a).However, here, we go beyond these reports by elucidating the electrophysiologi-cal mechanism behind this interaction: theta oscillatory coupling. In general, manyneocortical regions synchronize with hippocampal theta oscillations (Canolty et al.,2006). However, here, we demonstrated that specifically the mPFC exhibits in-creased coupling during memory integration. Thereby, we provide evidence fortheta-mediated functional interactions between these two key brain regions. Func-tional communication between the hippocampus and mPFC during memory inte-gration is supported by strong reciprocal anatomical connections. The anterior hip-pocampus has monosynaptic projections to the mPFC (Jay and Witter, 1991). In turn,the mPFC projects back to the hippocampus via the entorhinal cortex, in additionto a subcortical pathway with a thalamic relay (Xu and Südhof, 2013). These pro-jections from the mPFC to the hippocampus have recently been shown to play acrucial role in retrieving sparse hippocampal memory representations (Rajasethu-pathy et al., 2015) and are therefore important for memory integration throughretrieval-mediated learning. In addition, theta peak frequency has been found tocorrelate with structural connectivity between the hippocampus and mPFC, sug-gesting that theta oscillations are mediating interregional communication (Cohen,2011). But how might theta oscillatory coupling facilitate hippocampal-prefrontalneuronal interactions in service of memory integration? Oscillatory coupling hasbeen put forward as a mechanism for long-range information exchange betweenbrain regions (Fell and Axmacher, 2011). By synchronizing the excitable phases ofneuronal populations in distant brain regions, a window for effective communica-tion is established. Potentially, the hippocampus imposes phase-locking of neuronsin the mPFC, enforcing that only task-relevant inputs are selected and amplified ineach subsequent theta cycle. Alternatively, the mPFC might bias reconfigurationof hippocampal cell assemblies by entraining theta oscillations. Theta-dependentspatially selective hippocampal place cells are known to remap when encoding sim-ilar environments (Jezek et al., 2011). One could speculate that when encoding anew but similar memory, cells coding for the already existing memory need to bereconfigured (i.e., remapped) for successful integration. This reconfiguration pro-cess may be facilitated by resetting the phase of ongoing hippocampal theta oscilla-tions (Monaco et al., 2011), allowing the encoding of a novel combined memory. Inaddition, phase coupling between the hippocampus and mPFC may also enable ex-change of information represented by phase-coded neuronal population sequences(Jones and Wilson, 2005). Taken together, our findings are in line with the idea thattheta coupling provides the electrophysiological mechanism through which these

4

80

key regions interact and integrate novel information with an overlapping existingmemory.The hippocampus and the mPFC have been put forward as core nodes of the neuralcircuit for memory integration and generalization (Preston and Eichenbaum, 2013;Xu and Südhof, 2013). But do the two regions have specialized roles during memoryintegration? Computational models (Kumaran and McClelland, 2012) propose thatthe hippocampus encodes and retrieves specific associations, whereas the cortexextracts common features across events. Accordingly, the hippocampus separatesneural patterns associated with distinct events, whereas the mPFC might combinepatterns of overlapping events (Jo et al., 2007). Evidence from human neuroimagingstudies supports the pattern-separation function of the hippocampus, by demon-strating its involvement in various episodic memory tasks (Shohamy and Wagner,2008). Likewise, the mPFC has been implicated in generating adaptive responsesto current events based on past experience (Stokes et al., 2013). By accumulat-ing contextual information of overlapping episodic memories, the mPFC constructsmnemonic schemas or networks, which represent prior knowledge to guide decisionmaking (Preston and Eichenbaum, 2013; Tse et al., 2007). However, it remains unclearwhether the division of labor between the hippocampus andmPFC is strictly dichoto-mous, since both pattern completion and pattern separation are known to take placein the hippocampus. Hippocampal cells express firing patterns for overlapping con-texts, suggesting the hippocampus itself is also involved in generalization acrossepisodes (Eichenbaum et al., 1999). In addition, recent neuroimaging findings corrob-orate the idea that the hippocampus simultaneously performs episode segregationand integration (Milivojevic et al., 2015). Nonetheless, our results indicate that boththe hippocampus and the mPFC play an important role during memory integration,potentially via retrieval-mediated learning and pattern completion of overlappingmemories.Memory integration is the key process underlying regularity extraction and general-ization across similar events and situations. However, a tradeoff between memoryspecificity and generalization is vital to prevent maladaptive overgeneralization ofmemories. Here, we provide evidence for a crucial role of hippocampal-prefrontaltheta coupling in memory generalization. Further investigations of this electrophys-iological signature might improve our understanding of psychopathologies linkedto overgeneralization, such as posttraumatic stress disorder and depression (Xu andSüdhof, 2013). Moreover, our findings might guide future attempts to bias memoryintegration by manipulating or entraining region-specific theta oscillations. Facili-tating or impeding the integration of specific pieces of information might help us topotentially accelerate learning and enhance knowledge acquisition.Taken together, our findings highlight the involvement of the hippocampus andmPFC inmemory integration. Theta oscillations orchestrate the integration of mem-

4

81

ories by engaging the hippocampus and facilitating communication between thehippocampus and mPFC. These mechanisms constitute the crucial first step in theformation of relational memory networks, enabling us to assimilate information andultimately expand our knowledge base.

4.4.Methods4.4.1.Data AcquisitionParticipants performed an adapted version of the associative inference task used by Zei-thamova and Preston (Zeithamova and Preston, 2010) (Figure 4.1) while MEG data wererecorded (see subsection 4.5.2 for details). Experimental procedures were reviewed andapproved by the local ethical review committee (CMO committee on Research InvolvingHumans, region Arnhem-Nijmegen, the Netherlands). We randomly paired object stimuli tocreate 96 triad associations (ABC) and 48 dyads (YX). Participants were exposed to premiseassociations (AB and CB pairs) and control associations (YX pairs), followed by a memorytest in 12 independent cycles. Crucially, the AC association of a triad was never directlyencoded, although memory for this inferred association was tested. Each cycle comprisedtwo separate encoding blocks, followed by a test block, allowing us to assess memory per-formance. After an initial analysis of behavioral data (Figure 4.2A; see subsection 4.5.2 fordetails), seven participants were excluded based on their low inference performance level(criterion at double chance level: at least 50% correct, to ensure sufficient trials per con-dition). The MEG data of 20 high-performing participants in total were preprocessed (seesubsection 4.5.2 for details) and further analyzed.

4.4.2. Subsequent Integration ContrastTo isolate the neural oscillatory signatures of memory integration, we contrasted encoding-related activity during fully successful integration trials in block 2 (AB, CB, and AC cor-rect) with non-integration trials (AB and CB correct, CB correct, and YX correct). Cru-cially, a premise or direct association was nonetheless successfully encoded during all non-integration trials (Figure 4.3A). Thereby, we isolated activity related to successful AC infer-ence and subsequent integration into the ABC triad. To prevent bias in source activity esti-mation, we equalized the number of trials in each condition set to match the smaller subsetsize, by selecting a random subsample once. Across participants, on average 41 trials percondition entered the final analysis (range: 25-56 trials, SD: eight trials).

4.4.3. Source ReconstructionWith a strong a priori hypothesis on the hippocampus—a well-defined anatomical brainregion—we employed an ROI source reconstruction technique (Figure 4.5), where we cre-ated leadfields based on anatomical priors (Limpiti et al., 2006). Hereby, we aimed to com-pute one leadfield generated by the entire hippocampus, in contrast to the more traditionalapproach where one independently reconstructs a collection of point sources and averagesafterward. First, we spawned a regular 5-mm grid covering all voxels inside the ”Hippocam-pusL” and ”HippocampusR” anatomical masks from the Automated Anatomical Labeling at-las, with 2 mm smoothing, in Montreal Neurological Institute (MNI) space. Next, for eachparticipant, we normalized the MNI grid based on the participant’s brain morphology taken

4

82

from an individual structural MRI (see subsection 4.5.2 for details), so that each grid pointwould cover roughly the same anatomical location across par- ticipants. The brain tissue seg-ment from the structural MRI was used to construct a volume conduction model, based onthe single-shell method (Nolte, 2003). Using this model, we computed how a dipolar sourceat each grid point would project to the sensors, yielding a forward model in the form of asensors-by-grid point leadfield matrix (Figure 4.5, bottom). In a next step, we used singularvalue decomposition to reduce the number of columns in the leadfield matrix, by select-ing the top left-singular vectors explaining at least 95% of the variance. Each hippocampalROI leadfield matrix comprised six to eight spatial components. For the subsequent spa-tial filter estimation, we took the equalized sets of trials in each condition and combinedthem into one dataset. By using a balanced common filter approach, we aimed to preventa potential bias toward one of the conditions. Next, we applied a Fourier transformationto the data from the full 0- to 4,000-ms encoding window, using multitapering. 15 tapersfrom discrete prolate spheroidal sequences (DPSS) were used for spatial filter estimationwith 2 Hz spectral smoothing. From the complex-valued Fourier coefficients, we computedthe cross-spectrum (Figure 4.5, top) for our frequency bands-of-interest (see next section forspecifications). We used the entire encoding window—a continuous interval without visualstimulation—to improve estimation of the cross-spectrum. Next, we employed a DynamicImaging of Coherent Sources (DICS) beamformer (Gross et al., 2001) to estimate oscillatoryactivity at the source level. The cross-spectrum was regularized prior to matrix inversionby loading the diagonal of the matrix with 5% of the average sensor power. We used theDICS beamformer to fit a dipole for each of the spatial components and obtained a spatialfilter for each ROI (Figure 4.5, right). Subsequently, we projected Fourier-transformed singletrial sensor data through the spatial filter to reconstruct the source components comprisingeach ROI. To obtain theta power of the ROI as a whole, we combined information from eachsource component by taking the trace of the source cross-spectral density matrix. For thewhole-brain source reconstruction analysis, we employed a standard 8-mmMNI grid. Here,we projected the three resulting dipolemoments (x, y, and z direction) by taking the principaleigenvector of the real part of the cross-spectral density matrix (kept constant across trials).For the connectivity analysis, this projection method was also applied to obtain complex-valued Fourier coefficients for the left hippocampal ROI.

4.4.4. Theta Power AnalysisIn an initial step, we targeted the 3- to 7-Hz frequency band by using 2 Hz spectral smooth-ing centered on 5 Hz, with a 1,000-ms sliding time window in steps of 50 ms spanning a timewindow-of-interest from 0 to 2,000 ms. Spectral data from the three resulting orthogonalSlepian tapers were projected through precomputed spatial filters for the left and right hip-pocampus. We quantified differences between the integration and non-integration condi-tions by computing T-statistics of this contrast across participants. We tested for exchange-ability across conditions based on the resulting variance-normalized theta difference timecourse for the left and right hippocampus together, using a one-tailed, paired t-test (cluster-based permutation) with 100,000 permutations (time point cluster-inclusion criterion: p <0.05 nonparametric on individual time point level, cluster statistic: summed T-values). Fordisplay purposes, the theta difference time course was smoothed using shape-preservingpiecewise cubic interpolation. Power values from the peak time point showing the strongest

4

83

data sensor positions cross-spectral density

structural MRI segmentation

volume

conduction

model

anatomical

maskgrid lead!eld matrix reduction

region-of-interest

spatial !lter

sensors×

sensors

sensors×

grid

sensors×

time

Figure 4.5: Hippocampus-BasedMEG Source Reconstruction Procedure.Based on participant-specific anatomy, we constructed a realistic volume conduction model (middle).In parallel, we created a high-resolution grid spanning a specific anatomical ROI, aligned to a commontemplate space (bottom). Using the volume model and sensor position information, we computeda leadfield for each grid point and performed feature reduction on the resulting matrix (i.e., forwardsolution). A beamformer algorithm was used to compute a spatial filter, with the reduced leadfieldmatrix and data covariance structure (cross-spectral density) as input (i.e., inverse solution).

normalized difference were extracted for each individual condition, and the associated sig-nificance value of the differencewas obtained using a one-tailed nonparametric paired t-testwith 100,000 permutations. In addition, Bayes factors were computed using the standard-ized implementation of the Bayesian paired samples t-test in the JASP software packagev.0.7.1.12, to indicate how much more likely our hypothesis (i.e., more theta power in thesuccessful integration condition) is than the null hypothesis (i.e., no difference). For thefrequency-resolved follow-up analysis, we used a 1,000-ms sliding time window to cover the500- to 2,500-ms interval with steps of 100 ms. We explored frequencies from 2 to 12 Hz insteps of 1 Hz, with 2 Hz spectral smoothing. We applied the subsequent integration contrastto obtain T-value difference maps. The resulting time-frequency representations from theleft and right hippocampus were interpolated for display purposes. To obtain a whole-brainspatial distribution of the subsequent integration effect, we computed source activity in thefull 8-mm grid at the peak time point. We used a whole-brain cluster-based permutationpaired t-test (10,000 permutations, cluster statistic: summed T-values). The voxel clusterinclusion criterion was set to p < 0.01 (nonparametric on individual voxel level) in order toobtain separate statistics for left and right hemisphere clus- ters. For display purposes, weinterpolated the resulting maps to the MNI152 anatomical template with a resolution of 0.5mm and thresholded the maps at the cluster inclusion threshold value. All brain images are

4

84

displayed according to neurological convention.

4.4.5. Coupling AnalysisFor the seed-based functional connectivity analysis, we collected the complex Fourier outputfor both the left hippocampal ROI and the whole-brain grid at the peak time point revealedby the power analysis (1-s time window from -100 to 900 ms, 5 Hz center frequency with2 Hz spectral smoothing). Next, we computed across-trial coherence between the left hip-pocampus and each individual grid point, resulting in a whole-brain coherence map for eachparticipant. After Fisher-Z transformation of the coherence measure, we debiased the databy dividing by the square root of the summed inverse degrees of freedom in each condi-tion. The resulting debiased maps were subjected to a one-tailed cluster-based permutationpaired t-test across participants (10,000 permutations, cluster statistic: summed T-values)with a voxel cluster inclusion criterion of p < 0.01 (nonparametric on individual voxel level).Since we had a strong a priori hypothesis about the approximate brain region communicat-ing with the hippocampus, we restricted the statistical analysis to the anatomically delin-eated the mPFC. We used a hand-drawn mPFC mask from a previous fMRI memory integra-tion study, which encompassed all cytoarchitectonic subdivisions of the mPFC associatedwith the limbic system (Schlichting et al., 2015). We did not employ the ROI source recon-struction technique for the mPFC due to its extent and functional subparcellation but usedthe regular point source grid for the connectivity analysis instead. The mPFC mask in MNIspace was interpolated to this 8-mm grid space using nearest-neighbor interpolation. Posthoc statistics on the peak coherence voxel were obtained using a one-tailed, nonparametric,paired t-test with 100,000 permutations.

4.5. Supplemental Information4.5.1. Supplemental Figures and Tables4.5.2. Supplemental MethodsParticipantsThirty-seven healthy volunteers (21 female, age range: 18-31 years, average age: 22 years,SD: 3 years) without brain abnormalities and psychiatric or neurological history participatedafter giving their informed consent. All participants had normal or corrected-to-normal vi-sion and were reimbursed for their efforts. Experimental procedures were reviewed andapproved by the local ethical review committee (CMO committee on Research InvolvingHumans, region Arnhem-Nijmegen, the Netherlands). We excluded incomplete data setsfrom ten participants due to technical difficulties with the MEG equipment. Another sevenparticipants were excluded after initial behavioral analysis, based on their low inference per-formance level (criterion at double chance level: at least 50% correct, to ensure sufficienttrials per condition). MEG data of the remaining twenty high-performing participants (10female) were further analyzed.

Stimulus materialWeused 384 grayscale images of easily identifiable tools and utensils taken from theHemeraPhoto-Objects database and the Bank of Standardized Stimuli (Brodeur et al., 2010). Allobjects were centered, cropped and rescaled to 150 by 150 pixels. We normalized image

4

85

-3

T

3

0 22

12

fre

qu

en

cy (

Hz)

time (s)

Figure 4.6: Related to Figure 4.3. Right hippocampal theta power difference for subsequentmemory integration.Time-frequency representation of right hippocampal region-of-interest. Similar to the left hippocam-pus, an increase in theta power (5 Hz) at around 600 ms can be seen, albeit not significant.

Table 4.1: Related to Figure 4.3. Brain regions showing theta increased amplitude duringmem-ory integration.Besides bilateral temporal lobe, we observed clusters in occipital and brainstem regions. Occipital re-gions presumably play a role in task stimulus representation. Brainstem activation might be explainedby the nearby ventral tegmental area, a key brain region of the reward circuit and part of a theta-synchronized network comprising prefrontal and hippocampal regions (Fujisawa and Buzáski, 2011).Effects are thresholded at p < 0.01 to match Figure 4.3E; time point: 400 ms into the encoding inter-val (-100 to 900 ms window), cluster peak T-value corresponds to the specified MNI coordinates inmillimeters, cluster size is given in number of contiguous 8 mm grid points (voxels) showing an effectafter thresholding.

Anatomical region x y z T-value Clustersize

left middle temporal gyrus -76 -24 -16 3.92 143right superior temporal gyrus 44 -16 -8 4.07 104brainstem -4 -32 -40 3.23 18left middle occipital lobe -36 -64 8 3.20 6right postcentral gyrus 68 -8 16 2.39 1left superior frontal gyrus -28 24 32 2.67 1

luminance by pixel-wise adjusting the mean and standard deviation of the intensity values,using the SHINE toolbox for MATLAB. Stimuli were presented using Presentation software(v16.4, Neurobehavioural systems).

4

86

2

12

fre

qu

en

cy (

Hz)

-3

T

3

0 22

12

fre

qu

en

cy (

Hz)

time (s)

A

B

Figure4.7: Related toFigure4.3. LinearlyConstrainedMinimumVariancebeamformer results.Time-frequency representation of (A) left and (B) right hippocampal theta power difference for sub-sequent integration, using an alternative inverse solution method (see subsection 4.5.2). In the lefthippocampus, a positive early theta cluster (5 Hz) can be seen at around 600 ms.

Experimental taskWe adapted the associative inference task used by Zeithamova and Preston (2010) for ourpresent MEG experiment. Participants learned 96 triad associations (ABC) and 48 dyads(YX) of object stimuli. Premise associations (AB and CB pairs) and control associations (YXpairs) were each shown once, followed by a memory test. In total, participants completed 12independent cycles, each covering eight triads and four dyads each. To dissuade participantsfrom making excessive (magnetically-disturbing) eye movements when viewing the stimuli,we opted for brief sequential stimulus presentations, in contrast to parallel on-screen pre-sentation used in earlier fMRI studies (Zeithamova and Preston, 2010). Each encoding blocktrial commenced with a stimulus pair (two members of a triad or a dyad) being presentedsequentially against a gray background, within angle of five degrees from the center, an on-screen duration of 200ms each, and a 50ms blank screen interval in between the two stimuli.The offset of the second stimulus marked the onset of an encoding interval, during whichthe participant was allowed to rehearse the presented pair. During this period, only a whitefixation cross was on-screen. After 4000ms, the fixation cross turned red for 200ms duringwhich the participants were encouraged to blink and to prepare for the upcoming new trial.

4

87

A B

C

2

12

fre

qu

en

cy (

Hz)

0 22

12

fre

qu

en

cy (

Hz)

time (s)

2-2

T

L

Figure 4.8: Related to Figure 4.3. Sensor-level representations of theta power effects.(A) Topographical representation of theta power difference for subsequent integration in the peaktime-frequency window of the left hippocampal theta power effect (400 ± 500 ms, 5 ± 2 Hz). Time-frequency representations averaged across (B) left and (C) right temporal sensors shown as highlighteddots in the topography in (A).

Each trial concluded with a jittered inter-trial interval between 2,000 and 2,500 ms with awhite fixation cross on-screen. During the first encoding block, participants learned the ABassociation of the triads (A as first stimulus and B as second stimulus) and were exposedto an XX control condition (twice the × stimulus). During the second encoding block, theCB association was presented (C as first stimulus and B as second stimulus), together withthe YX control condition. The specific order of presentation was chosen to ensure that anypotential retrieval of previously learned information ensuing cue presentation would occurwith comparable delays. During the test block, a retrieval cue was presented for 300 ms,followed by a 1,500 ms retrieval phase with only a fixation cross. The cue stimulus couldbe any triad or dyad member. After the delay, participants selected the associated triador dyad member from four alternatives (one correct and three randomly drawn from theencoded objects in that block) by pressing a corresponding button on a MEG-compatiblebutton-box with their right hand. The four alternatives were simultaneously on-screen in arow for 3000 ms. Participants could respond for another 2,000 ms after the stimuli disap-peared. The moment the participant responded with a button-press or in case no response

4

88

0 2

10

fre

qu

en

cy (

Hz)

time (s)

90

A B

R

2

T

-2

Figure 4.9: Related to Figure 4.3. Sensor-level subsequent integration effects in high fre-quency bands.(A) Time-frequency representation averaged across all MEG sensors. We observed a late significantbeta power decrease (p < 0.002 cluster-corrected, delineated with dashed white line). There were noother significant differences (p > 0.21 cluster-corrected, see subsection 4.5.2 for details). (B) Topograph-ical representation of the beta power decrease, averaged across the cluster time-frequency window(12-29 Hz, 800-2,500 ms, see outline in A). The observed beta power decrease accords with recentframeworks implicating low frequency power decreases in subsequentmemory paradigms (Hanslmayret al., 2012).

was given within 5,000 ms, a second response screen followed during which participantsattached a confidence rating to their choice by again pressing one of four response buttons.Confidence was measured on a scale from one to four, ranging from “just guessing”, “lessconfident”, “more confident” to “(almost) sure”. Immediately after a response or maximally3,000 ms, a jittered intertrial interval of 1,500-2,000 ms preceded the next trial. Crucially,memory for the AC association of any given triad was always probed before AB or CB asso-ciations, to prevent learning during the test block. Participants were instructed to activelyencode the associations during the rehearsal period and make the AC link during the CBencoding trial. No further elaborations on encoding strategy were provided. In addition,we requested participants to indicate the lowest confidence rating in case they consciouslymade their response decision by excluding the other three of four alternatives. We prepared

4

89

0

T

3

A

½π

-½π

π 0

½π

-½π

0π

B

L R

Figure 4.10: Related to Figure 4.4. Theta phase coupling duringmemory integration.(A) Brain regions showing a higher theta phase-locking value with the left hippocampus during subse-quent integration. A similar pattern of results with a cluster in the mPFC (peak: x, y, z = [4, 48, 0] T₁₉= 2.69) was observed as in the seed-based coherence analysis reported in the main test. Slices werecentered on the coherence peak in the mPFC from Figure 4.4C (x, y, z = [-4, 40, -8]) to allow compari-son. Maps were thresholded at p < 0.05 for display purposes. (B) Theta phase lags between the mPFCcoherence cluster peak and the left hippocampus in the integration and non-integration conditions,where each line represents one participant. In case coupling is driven by volume conduction of sig-nal from a single source, we would observe a clustering of phase delays around zero. However, wefound no evidence for a non-uniform distribution of phase lags across participants (Rao’s spacing test(Batschelet, 1981), integration condition: U = 122, p > 0.5; non-integration condition: U = 143, p > 0.5).The absence of zero phase lag clustering suggests that the observed mPFC coupling effects are notdue to volume conduction.

participants with a written instruction text, followed by a pre-training practice cycle outsidethe MEG system and additional reminder examples prior to the start of data acquisition.Stimuli used for training were not used for the actual experiment.

Behavioral analysisFor all but one of the participants, results from all 12 cycles entered the analysis, with a totalof 336 test trials (one participant only completed 10 cycles due to technical difficulties). Wecomputed the percentage correct responses for each different association type and testedfor significant differences across the full group of 27 participants, with six two-tailed pairedt-tests. To account for non-normality in the data due to potential ceiling effects, we usednonparametric permutation tests with 100,000 permutations to obtain significance values.We applied a Bonferroni-correction for multiple comparisons.

4

90

Table 4.2: Related to Figure 4.4. Brain regions showing increased theta coupling with the lefthippocampus duringmemory integration.Besidesmedial prefrontal regions, several other regions exhibited increase coupling. Notably, the righthippocampus showed some evidence for theta coupling with the left hippocampus during successfulintegration, suggesting interhemispheric functional interactions. In addition, a large cerebellar clusterwas observed, extending into right hemisphere higher-order ventral visual brain regions, likely involvedin task stimulus representation and reactivation (Staresina et al., 2013). Effects are thresholded at p< 0.01 to match Figure 4.3E; time point: 400 ms into the encoding interval (-100 to 900 ms window),cluster peak T-value corresponds to the specified MNI coordinates in millimeters, cluster size is givenin number of contiguous 8 mm grid points (voxels) showing an effect after thresholding.

Anatomical region x y z T-value Clustersize

right cerebellum crus 1 36 -56 -40 4.78 60left inferior temporal gyrus -76 -24 -32 4.15 14right hippocampus 20 -8 -16 3.85 14left middle frontal gyrus, orbital -4 40 -8 2.97 14left inferior frontal gyrus, opercular -60 16 32 3.14 4right inferior temporal gyrus 76 -40 -24 2.97 4left inferior parietal gyrus -68 -40 40 3.54 3white matter -28 -40 24 2.67 2left angular gyrus -44 -56 24 2.61 1left superior temporal gyrus -68 -48 16 2.67 1left thalamus -12 -16 8 2.67 1right inferior occipital lobe 28 -96 -8 2.60 1left cerebellum 8 -28 -56 -40 2.45 1

MEG data acquisition and preprocessingWe used a whole-head 275-channel axial gradiometer MEG system (VSMMedTech Ltd., CTFSystems, Coquitlam, CB, Canada), located in a magnetically shielded room. Due to two mal-functioning channels, we acquired data from 273 sensors. Participants were seated 80 cmbehind a screen on which we back-projected task material with a projector. The MEG signalwas low-pass filtered at 300 Hz prior to digitization and recorded with a 1200 Hz samplingrate. Head position relative to the gradiometer array was monitored using localizer coils at-tached to the participant’s nasion and ear canals, and kept stable using custom online tools(Stolk et al., 2013) with a maximum displacement of 5 mm from the starting position. Incase the movement criterion was exceeded, we readjusted the participant’s head positionin the breaks between study-test cycles. In addition, we used high-resolution eye-tracking(SR Research Eyelink 1000) to monitor eye movements during the task. Data were analyzedusing MATLAB (The MathWorks, v2014a) with the FieldTrip Toolbox (Oostenveld et al., 2011)(v20121231 for preprocessing, v20150601 for source analysis). We epoched the continuousrecording from 1500 ms before each encoding interval to 4000 ms after the start of theinterval (5500 ms total length). SQUID sensor jumps and muscle artifacts were detectedby Z-scoring and aggregating the appropriately processed MEG signals, using FieldTrip’s de-fault preprocessing settings for each artifact type. Subsequently, the Z-scores were visually

4

91

inspected and epochs with unusually high score were excluded from subsequent analysis. Ina next step, we high-pass filtered the remaining epochs (4th order Butterworth Infinite Im-pulse Response filter, 0.5 Hz cutoff) and computed 3rd order synthetic gradients to furtherremove ambient noise. The logistic infomax Independent Component Analysis algorithmwas employed to unmix theMEG sensor data into independent temporal components, fromwhich putative heartbeat, horizontal- and vertical eyemovements and eye blink componentswere visually identified and subsequently removed from the data. Finally, any remainingepochs with unusual high variance (Z-score) were manually removed, before downsamplingthe remaining epochs to 600 Hz. To improve source reconstruction accuracy, we obtainedindividual participant T1 anatomical scans using a 1.5T MR scanner (MAGNETOM Avanto;Siemens Healthcare) with an MPRAGE sequence with 1 mm isotropic voxels. Images weremanually aligned to the averageMEGhead position using the nasion and ear canal referencepoints, and segmented to delineate the brain tissue and the inner surface of the skull.

Sensor-level analysisFor the sensor-level analysis, we computed the synthetic planar gradient, followed by a time-resolved frequency domain transformation identical to the source-level analysis (1,000 mssliding window, -500 ms to 2500 ms in steps of 100 ms, 2 to 90 Hz in steps of 1 Hz, 2 Hzspectral smoothing). We obtained the topographical sensor-level distribution of differencevalues corresponding to our temporal hippocampal theta peak and, based on our temporallobe hypothesis, we averaged the accompanying time-frequency representations (interpo-lated for display purposes), for left and right temporal sensors (Figure 4.8). In an additionalexploratory analysis, we computed power differences across time for a broader range of fre-quencies (8-90 Hz). On these sensor-level data, we performed a statistical analysis using atwo-tailed paired T-test (cluster-based permutation) (Maris et al., 2007) with 5,000 permu-tations (cluster inclusion criterion: p < 0.05 parametric, cluster statistic: summed T-values).

Eye-movementsIn addition to removing artifact components from theMEG data using ICA, we inspected theeye-tracker signal for potential differences between conditions. To quantify the amount ofeye movement-related activity, we combined data from the two eye-tracker channels fromthe 0 to 4000 ms encoding interval and extracted the square root of the pooled variancesacross trials of each condition. We observed no significant differences between conditions(T₁₉ = 0.94, p = 0.40, nonparametric paired samples t-test with 100,000 permutations, two-sided) with minor evidence supporting equal eye-related variance in both conditions (BF₁₀ =0.34, support for null-hypothesis: BF₀₁ = 2.9).

Linearly Constrained Minimum Variance beamformer analysisTo corroborate our DICS beamformer theta power results with an alternative source recon-struction method, we employed the Linearly Constrained Minimum Variance (LCMV) beam-former (Veen et al., 1997). The LCMV beamformer operates in the time domain instead ofthe spectral domain, using the (typically spectrally broad-band) sensor covariance matrix,rather than the spectrally-confined cross-spectral density matrix for the computation of thespatial filters. As a consequence, the Fourier transformation takes place after reconstructingthe source signals and not prior to beamforming. This possibly results in a smoother time-frequency representation, but may lead to detection sensitivity issues under low signal-to-

4

92

noise conditions. Settings of the LCMV beamformer were kept as similar as possible to theDICS beamforming procedure. We bandpass filtered the full epochs (2 to 12 Hz, windowed-sinc finite impulse response filter) and projected single trial data through precomputed spa-tial filters (0 to 4,000 ms window) and Fourier transformed the resulting virtual channeldata (2 to 12 Hz in steps of 1 Hz with 2 Hz spectral smoothing, 1,000 ms sliding time windowcovering the -500 ms to 2,500 ms interval in steps of 100 ms). T-value difference maps wereobtained by applying the subsequent integration contrast. For display purposes, we interpo-lated the resulting time-frequency representations from the left and right hippocampus.

Phase-locking values and phase lag analysisTo investigate phase couplingwith an alternative connectivitymeasure, we computed phase-locking values between the left hippocampal ROI and the whole-brain grid from the complexFourier output (Lachaux et al., 1999). After fisher-Z transformation, we obtained T-statisticsof the subsequent integration contrast for comparison with the hippocampal-prefrontal co-herence effect (Figure 4.10A). We used nonparametric significance values (10,000 permuta-tions, individual voxel level) to threshold the final map. To investigate whether our effectswere potentially due to volume conduction of signal from a single source, we extractedmeanphase delays across trials between the left hippocampal ROI and mPFC peak voxel from thecoherence analysis. Any non-zero phase delay is not attributable to signal leakage, whereasa zero phase delay could potentially be due to volume conduction. We displayed thesemeanphase lag values for each condition and for each participant to confirm the absence of clus-tering around zero (Figure 4.10B) and used Rao’s Spacing Test (Circular Statistics Toolboxfor MATLAB) to statistically test non-uniformity of the phase lag distributions (Batschelet,1981).

5General discussion

95

5

96

The experimental work presented in this dissertation yielded several key observa-tions with impact on three core domains. Firstly, the insights from the experi-ments further our neuroscientific understanding of the neural mechanisms support-ing associative memory and pave the way for future research. Secondly, significantmethodological advancements over the course of these experiments. The innova-tive methods presented here provide neuroscientists with new tools to investigatethe brain. Finally, the techniques and insights from these experiments can poten-tially be applied for clinical diagnostics and in educational settings. In the followingsection, I will embed the results from the experimental chapters in these three do-mains.

5.1.Neuroscientific insightsIn this thesis, I set out to answer the following two-part neuroscientific question:how does the brain connect memories and how do memories connect the brain?By looking at the motif of convergence in Chapter 2, we approached the two partsof this question simultaneously in one experiment. Computational models have es-poused the importance of the hippocampus as convergence zone, binding differentaspects of an episode into a coherent conjunctive representation by integrating in-formation frommultiple brain regions (Marr, 1971; Damasio, 1989; McClelland, 1994).However, evidence for this long-held hypothesis is limited, since previous work haslargely focused on representational and network properties of the hippocampus inisolation (Moita et al., 2003; Chadwick et al., 2010; Vincent et al., 2008). Here, weleveraged a combination of representational and connectivity analysis. First, we em-ployedmultivariate pattern analysis on functional magnetic resonance imaging datafrom an associative memory task. We combined this representational analysis withgraph-theoretical network analyses, in order to test the idea that the hippocampusacts as mnemonic convergence zone. The representational component of our ap-proach enabled us to investigate where in the brain memories, or separate elementsof a single episodic memory, are connected and how they are combined and storedas a conjunctive memory representation. Complementarily, the network compo-nent of our approach was aimed at quantifying the amount by which memories -specifically the retrieval of associative information - connect different brain regionsto each other. In line with our predictions, we observed a striking overlap of conjunc-tive coding and hub-like network attributes in the hippocampus. Thus, we providedcompelling evidence that the hippocampus acts as a mnemonic convergence zone.We were the first to apply both complementary analyses and, in this manner, devisea tailor-made method to test the hypothesis that the hippocampus acts as a conver-gence zone. In Chapter 3, we further investigated the nature of conjunctive represen-tations by looking at the emergence of associative memories. Here, we attemptedto map the network geometry - how the brain connects memories - of newly ac-

5

97

quired associative memories, by comparing hippocampal activity patterns beforeand after learning. We observed increased neural pattern similarity between associ-ated items, compared to non-associated items, as a consequence of learning. Thus,we showed that representations in the brain becomemore similar - more connected.This finding is in line with a series of previous fMRI studies that compared neural pat-tern similarity before and after learning, using various memory tasks. Group-levelchanges in pattern similarity have been reported in fear conditioning (Visser et al.,2011), implicit temporal regularity extraction (Schapiro et al., 2012, 2013, 2015), transi-tive inference (Schlichting et al., 2015) and narrative insight tasks (Collin et al., 2015;Milivojevic et al., 2015). These reconfigurations have been observed in several brainregions, such as the insular cortex (Schapiro et al., 2013) and mPFC (Milivojevic et al.,2015), but also in the hippocampus (Collin et al., 2015; Milivojevic et al., 2015). Here,we complemented this list of more complex memory paradigms with basic explicitpaired-associate learning. We provided important insights into the neural mecha-nisms involved in this well-understood keystone memory task and neuroscientificmodel of episodic memory: by tracking the emergence of associative memories, wewere able elucidate how, where and when these representations are formed. Firstly,we showed that the formation of associativememories entails changes in neural pat-tern similarity, making associated items more similar. Possibly, these changes in thecross-correlations of multi-voxel fMRI activity patterns are related to the formationor modification of engram complexes, representing the paired items in conjunction.This idea accords with electrophysiological recording studies showing that cell as-semblies adapt their stimulus-specific firing patterns as a function of learning, andare thus tuned to process conjunctive information (Sakai and Miyashita, 1991; Isonet al., 2015). Secondly, we showed that the hippocampus houses these newly-formedassociative representations. This observation is in line with our findings from Chap-ter 2, where we identified the hippocampus asmnemonic convergence zone. Thirdly,by interrogating hippocampal representations in a neutral task setting before and af-ter learning, we demonstrated that the neural reconfiguration process takes place onthe short term, as a consequence of learning. This finding complements the observa-tions reported in Chapter 2, where we investigated associative representations dur-ing memory retrieval. Taken together, the set of representational analysis presentedin both chapters provide compelling experimental evidence for the long-held conjec-ture of computational models that associative representations underlying episodicmemories are located in the hippocampus (Marr, 1971). Furthermore, this conclusionis supported by the existence of specialized cell types in the hippocampus, that codefor spatial location (O’Keefe and Dostrovsky, 1971), remembered objects (Manns andEichenbaum, 2009) or conceptual information (Quiroga et al., 2005). These celltypes are thought to encode episodic memory representations, by mapping the con-junctions between features of an episode (Eichenbaum, 2000). In addition, this idea

5

98

accords with the hippocampal index theory (Teyler and DiScenna, 1986), where con-junctive representations provide the pointers to the various neocortical represen-tations of these epsiode features. Our findings complement the substantial bodyof evidence from fMRI studies that support the presence of conjunctive associativerepresentations in the hippocampus (Chadwick et al., 2010; Shohamy and Wagner,2008; Staresina et al., 2013; Davachi et al., 2003; Kuhl et al., 2013; LaRocque, 2013; Azabet al., 2014; Copara, 2014; Rissman and Wagner, 2012; Milivojevic et al., 2015; Collinet al., 2015). Moreover, in Chapter 2, we provide additional evidence for the indextheory by relating hippocampal memory representations to a central hub-like roleduring memory retrieval. The crucial role of the hippocampus in connecting differ-ent subnetworks of the brain resonates with the idea that hippocampal representa-tions index other brain regions, such as sensory cortex. Taken together, the findingspresented in this thesis show that, both in terms of representations and connected-ness to the rest of the brain, the hippocampus qualifies as a mnemonic convergencezone. Building on this conclusion, we investigated the electrophysiological mecha-nisms of associative memories in the hippocampus and the integration of memoriesin particular. Integration of separate memories forms the basis of inferential reason-ing - an essential cognitive process that enables complex behavior. In Chapter 4, weusedMEG to study the brain oscillations supporting memory integration. Firstly, weshowed that amplitude of theta oscillations from hippocampal sources predicts suc-cessful integration of memories. This finding is in line with the long-established roleof theta in hippocampal function and its involvement in memory processes (Win-son, 1978; Buzsaki and Moser, 2013). In general, a theta increase indicates that thehippocampus is active and ready to encode new information and retrieve memories(Buzsáki, 2002). Our observation that theta oscillations are more prominent whenmemory integration is successful accords with the retrieval-mediated learning hy-pothesis (Shohamy andWagner, 2008). This hypothesis posits that a newmemory isincorporated into an existingmemory network, by retrieving previously storedmem-ories that share features or context, and subsequently re-encoding a unifiedmemory(Kumaran and McClelland, 2012). Here, both retrieval and encoding processes arerequired during integration, which might explain the boost of theta oscillations toengage the hippocampus. In addition, we studied how memory integration is sup-ported by brain connectivity. Specifically, we test the hypothesis that functionalinteractions between key brain regions are mediated by theta oscillations. Consid-erable evidence suggests cross-talk between the hippocampus and mPFC plays acrucial role inmemory integration (Preston and Eichenbaum, 2013; Zeithamova et al.,2012a; Schlichting et al., 2015). Although previous work indicates that theta oscilla-tions subserve these interactions during other memory and decision-making pro-cesses (Hyman et al., 2005; Siapas et al., 2005; Jones and Wilson, 2005; Young andMcNaughton, 2009; Benchenane et al., 2010; Sigurdsson et al., 2010; Fujisawa and

5

99

Buzáski, 2011; Miller, 2000; Kaplan et al., 2014; Garrido et al., 2015; Anderson et al.,2010), the electrophysiological mechanisms of cross-region communication duringmemory integration remained unknown. Here, we observed that increased thetacoherence between the hippocampus and mPFC during encoding predicts subse-quent memory integration. This finding is in line with our observations from Chap-ter 2, where we revealed the hub-like role of the hippocampus during associativememory retrieval. Taken together, these findings suggest that memories connectthe brain in a specific way: by connecting to different brain regions, the hippocam-pus plays a key role in brain communication by acting as a general network hubduring memory processes. In addition, more specific theta-mediated interactionswith the mPFC are crucial for successful memory integration. These conclusions arein line with studies implicating the mPFC in storing contextual representations ofmemories (Tse et al., 2007; Hyman et al., 2012) and thus suggest it plays an impor-tant role during the integration of memories with contextual overlap (Preston andEichenbaum, 2013). In addition, evidence suggest that the mPFC is involved in cog-nitive control during memory encoding and retrieval, specifically guiding retrievalprocesses and reconciling conflicting sources of information. Thereby, the mPFCmay provide a mechanism for dynamic switch between the old and new memorywhen both are integrated through their shared context. Additionally, hippocampaltheta oscillations have been put forward as an electrophysiological mechanism fordynamic switching between encoding and retrieval states (Hasselmo et al., 2002).Here, the alternating phases of theta determines whether information is stored orretrieved. Accordingly, the phase of theta has been found to be dependent on an-imal behavior (Hyman et al., 2003). Moreover, electrophysiological studies havefound that strength of inputs to CA1 and hippocampal synaptic plasticity dependon theta phase (Hyman et al., 2003). By separating encoding and retrieval, the mem-ory system may be able to avoid interference of previously encoded memories withsensory-related information (Hasselmo et al., 2002). By investigating the source-level oscillatory underpinnings of cross-episode memory integration in humans, weprovided the first evidence for the involvement of theta oscillations in this cogni-tive process, in accord with the retrieval-mediated learning hypothesis. All in all, werevealed an important electrophysiological mechanism underlying inferential rea-soning, memory-based decision-making and ultimately knowledge acquisition.In sum, the findings presented in this thesis shed light on the neural mechanisms ofassociative memory (Figure 5.1). Specifically, we elucidated the representational andconnectivity profile of the hippocampus with regard to associativememories. Firstly,the brain connects memories bymeans of conjunctive representations, stored in thehippocampus. These representations are formed bymodifying neural firing patterns,in away thatmultiple neurons respondmore similarly to two connected items, a phe-nomenon we were able to quantify using fMRI pattern analysis. Furthermore, theta

5

100

thetaoscillations

before learning after learning

connected memories

connected brainconnected brainconnected brainconnected brainconnected brainconnected brainconnected brainconnected brainconnected brainconnected brainconnected brainconnected brainconnected brain

Figure 5.1: How the brain connects memories and howmemories connect the brain.Summary graphic outlining the core neuroscientific insights. Isolated features or memories (left) areconnected through experience and learning (middle), and form a coherent episode or memory net-work (right). For example, imagine you meet three previously unrelated scientists on a dinner cruise.Subsequently, your memories of these scientists become related through the shared experience (recallFigure 1.1). In the brain, the hippocampus (green) represents the conjunctive representations, under-lying these memory networks and is highly connected with the rest of the brain during informationretrieval (converging arrows). Theta oscillations allow the hippocampus to communicate with the me-dial prefrontal cortex (yellow) and thereby integrate new memories into existing networks.

oscillations support connections between separate memory episodes. Secondly,memories connect the brain by linking the hippocampal index representations toneocortex regions when retrieving a memory, or establish targeted communicationwith the mPFC when integrating memories. These functional interactions are me-diated by coupled theta oscillations, synchronizing brain regions and providing amechanisms for cross-regional communication. Together, these key observationspave the way for further research on the mechanisms of associative memory (seeBox 7).

Box 7: Outstanding questions

In the current work, we showed that the hippocampus acts as a mnemonic conver-gence zone during the retrieval of associative memories. This finding accords withthe proposed indexing function of the hippocampus, although a direct relationshipremains to be demonstrated. What is the precise relation between hippocampal and

5

101

cortical memory representations? Recent work has shown that hippocampal activitycovaries with cortical memory reactivation (Bosch et al., 2014). But how is this rein-statement of information related to conjunctive representations in the hippocam-pus? Does reinstatement of content-specific information engage the memory hub-mode of the hippocampus? A second question is whether mnemonic convergencemetrics employed in Chapter 2 are stable over time, and how they might vary as afunction of normal or pathological aging. For instance, is the coincidence of hub-likeconnectivity and conjunctive information in the hippocampus degraded in patientswith Alzheimer’s disease? Furthermore, in Chapter 3, we demonstrated that associa-tive learning engenders a change in neural activity patterns. Wemight question howthis pattern-similarity change, which we measure on the fMRI voxel-level, is broughtabout. How do representations change from pre-learning to post-learning? Are theunrelated pre-learning representations replaced by a unified conjunctive, index-likerepresentation after learning? Or are representations connected by the shared partof their engram complexes, for instance by cell assemblies coding for the learningcontext? One potential way of investigating this unresolved issue is by comparingpre-learning with post-learning representations. However, although attempts havebeen made (Schapiro et al., 2012), the relatively long time interval between noise-sensitive scan sessions hinders fMRI multi-voxel pattern analysis. Another optionis to investigate the representational geometry of complex associative memory net-works, in addition to simple paired-associates. Are the magnitude and variance ofpattern similarity effects dependent on the complexity of the associative structure?The results of such a studymight provide clues on the nature of the pattern-similaritychange. Furthermore, we might wonder whether associative memory network re-configurations are stable over time, or rather vary as a function of memory consol-idation (Takashima et al., 2006; Frankland and Bontempi, 2005). Finally, it wouldbe interesting to see whether the paradigm extends from simple paired-associatesto abstract conceptual knowledge. For instance, does reading of encyclopedia con-cept descriptions, with a certain degree of textual overlap, elicit a similar change inneural representations? Can we track the acquisition of knowledge? The findingspresented in Chapter 4 have shown that theta oscillations provide a systems-levelmechanism for hippocampal-prefrontal interactions during memory integration andpotentially knowledge acquisition. But what are the cellular mechanisms of memoryintegration? Is memory integration linked to cell remapping, as seen in place cells(Colgin et al., 2008)? What is the role of hippocampal and medial prefrontal pat-tern completion and pattern separation processes in memory integration? In addi-tion, given the fact that theta oscillations are known to clock cyclic bursts of gammaband activity, we might wonder whether there is cross-frequency coupling involvedin memory integration, similar to other memory processes (Axmacher et al., 2010;Staudigl and Hanslmayr, 2013). Finally, although a previous fMRI study reported evi-dence for content-specific activations during memory integration (Zeithamova et al.,2012a), decoding of multivariate spectro-spatio-temporal activity patterns (Stokeset al., 2015) may provide additional insights into the timing and oscillatory mecha-nisms of retrieval-mediated learning. Further research is needed to address these

5

102

questions.

5.2.Methodological advancementsIn the experimental work presented in this thesis, we employed several innovativemethods in order to answer important neuroscientific questions. In Chapter 2, weused a novel combined approach of multivariate pattern analysis and graph theoryto simultaneously probe the representational and network properties of the brain.Hereby, we identified the human hippocampus as a mnemonic convergence zone,both in terms of conjunctive representations and its connectedness to the rest ofthe brain duringmemory retrieval. From a broader perspective, our approach equipsneuroscientists with a new tool to investigate mnemonic convergence in the brain.We devised a generic metric of convergence by quantifying the amount of overlapbetween conjunctive coding and hubness in a specified region. This metric maybe used in future studies, which aim to investigate mnemonic convergence acrossdifferent brain regions, during various cognitive task conditions, or across subpopu-lations.In Chapter 3, we investigated the formation of associative memory representationsin the brain using non-invasive neuroimaging techniques and a simple associativelearning paradigm. We used differential RSA where we looked at neural patternssimilarity before and after learning. Crucially, brain activity patterns were acquiredduring a independent task, unrelated to the learning phase, in order to obtain cleanerpatterns similarity estimates. We observed increased neural similarity between as-sociated items, compared to non-associated items, across a group of individuals.By demonstrating the ability to probe associative memory networks and learning-related reconfigurations of those networks in the brain, we validated differentialRSA as an important tool for memory research. This approach may be leveragedby future studies, which aim to investigate representational reconfigurations as afunction of different learning strategies, educational systems, types of material orthe stability of associative memory networks in clinical populations. In addition, weattempted to interrogate individual associative memory networks by reconstructingtheir representational geometry based on similaritymeasures between hippocampalactivity patterns. We assessed the predictive value of the reconstructed associativememory networks by classifying learned associations based on the representationalgeometries. However, this predictive value was found to be low, despite our currentstate-of-the-art image processing and analysis techniques. Nevertheless, our fMRIdataset and differential RSAmethods may serve as a benchmark for future attempts.Novel statistical approaches, for instance improved time series modeling (Mumford

5

103

et al., 2014) or probabilistic network reconstruction approaches (Hinne et al., 2015),benefit from the availability of a complete fMRI dataset acquired with the simplewell-understood associative learning paradigm. In addition, we outlined two com-plementary classification algorithms, which allow us to assess the predictive valueof reconstructed associative memory networks. Sharing part of our dataset andanalysis code with the community would enable other research groups to test newmethods and compare their performance against our benchmark results by cross-validating on left-out datasets (Freeman, 2015). Data sharing via competitive datascience platforms may significantly accelerate the development of improved indi-vidual associative network reconstruction methods.In Chapter 4, we applied a state-of-the-art innovative MEG source reconstructionmethod to investigate oscillatory signals from the hippocampus, for the first time.Hereby, we provided a practical demonstration of hypothesis-driven deep-sourceMEG and highlight the value of this approach for the human memory field. Conven-tionally, insights into the hippocampal mechanisms supporting memory functionscome from two separate fields: firstly, human fMRI studies have shown the roleof slow timescale hippocampal activation and content of hippocampal representa-tions in the healthy human brain, on the voxel population level. In contrast, invasiveintracranial electrophysiological recordings in animals and occasionally human pa-tients have elucidated the fast timescale hippocampal oscillatory mechanisms ofmemory functions. However, these electrophysiological recording studies can onlysample from a limited set of preselected brain regions-of-interest and are thereforeunable to demonstrate spatial specificity. By investigating the source-level oscil-latory underpinnings of cross-episode memory integration in healthy humans us-ing non-invasive MEG recordings and advanced source reconstruction methods, webridge the gap between invasive fast timescale recording studies on hippocampaltheta oscillations and slower BOLD activation effects revealed by human fMRI stud-ies. Our hypothesis-driven region-of-interest-based MEG source reconstruction ap-proachmay be used to further study the oscillatorymechanisms supportingmemoryfunctions.In sum, the experimental work presented in this thesis has produced three inno-vative methods that may be leveraged to investigate memory and other neurocog-nitive functions. Firstly, we proposed a generic metric of mnemonic convergence(Chapter 2). Secondly, we outline techniques to reconstruct the representationalgeometry of associative memory networks from fMRI data and ascertain their pre-dictive value on the level of an individual person (Chapter 3). Thirdly, we employadvanced deep-source reconstruction methods on MEG data to study hippocampaloscillatory signals (Chapter 4). Finally, these three methodological approaches maybe combined in future research projects (see Box 8).

5

104

Box 8: Composite approaches

Combining the methods developed and applied in this dissertation opens up newavenues for future research. Our generic metric of convergence (Chapter 2) maypotentially be combined with associative memory network reconstruction, usingdifferential RSA (Chapter 3). This conjoint approach would allow neuroscientiststo investigate how and when mnemonic information converges in the hippocam-pus and other brain regions. Secondly, we can investigate the electrophysiologi-cal signals of mnemonic convergence using MEG. In this thesis, we demonstratedthe feasibility of investigating frequency-specific connectivity changes by relatinghippocampal-prefrontal theta coupling to memory function (Chapter 4). In addition,previous studies have applied graph analysis to whole-brain source-level MEG dataand investigated the distribution of several hub measures (Hipp et al., 2012). Fur-themore, multivariate pattern analysis on MEG data has become increasing popu-lar (Stokes et al., 2015) and we have made successful attempts in decoding spectro-spatio-temporal activity patterns (van de Nieuwenhuijzen et al., 2013). In combina-tion, these methods may be leveraged to investigate the spectral and spatial overlapof conjunctive representations and hub shifts in network connectivity, on a millisec-ond timescale. Thirdly, successfully reconstructing the representational geometry ofassociative memory networks from electrophysiological data, such as MEG or EEG,would prove invaluable for potential brain-computer interface applications (van Ger-ven et al., 2009). Compared to fMRI, the relative high temporal resolution and porta-bility of electrophysiological recording equipmentwould permit responsive real-timefeedback.

5.3. ApplicationsThe experimental findings and innovative methods described in the previous sec-tions may be applied in several settings with significant societal impact. Firstly, find-ings may be used for clinical diagnostics. The generic metric of mnemonic conver-gence presented in Chapter 2 can be leveraged to quantify hippocampal integrityduring pathological aging. In this way, the method may be employed for early de-tection of conditions such as Alzheimer’s disease. Amyloid-beta deposition levels inpatients with Alzheimer’s disease have been found to peak in hub brain regions, suchas the hippocampus (Buckner et al., 2009) and degradation of these brain regionsis associated with impaired episodic memory function (Mielke et al., 2009). Ourgeneric metric of convergencemay be able to track progression of such degradation,related to memory function. In addition, the findings presented in Chapter 4 couldimprove our understanding of psychopathologies linked to memory overgeneraliza-tion. Memory integration lies at the basis of generalization across events. Certaindiseases, such as depression and posttraumatic stress disorder, have been linked

5

105

overgeneralization of memories. Therefore, being able to measure the electrophys-iological neural substrates of memory integration may provide a stepping stone to-wards a biomarker for these psychopathologies. In addition, disrupted hippocampal-prefrontal theta coherence been reported in mouse models of schizophrenia (Sig-urdsson et al., 2010). The deep-source oscillatory coupling methods employed inthe Chapter 4 may provide means to investigate this finding in human patients non-invasively and help to improve our understanding of this disease. In all, there areseveral anchor points for clinical applications of the work presented in this thesis.

after learningbefore learning

COMPARE

LEARNMEASURE

1

2

3

4

thetarepresentationalsimilarity analysis

canonical knowledge

Figure 5.2: Engineering knowledge with brain-based user models.In the example illustrated here, we suppose someone wants to learn facts about Samuel Eto’o, aCameroonian professional footballer who played for Chelsea FC. Initially, we probe the person’s brainrepresentations, using neuroimaging and RSA, and find that certain facts about football, England andthe footballer Klaas-Jan Huntelaar are present (red nodes, 1). We compare this neural knowledge net-work with a canonical knowledge representation based on big data (2). Next, we present the missingelements (green nodes) and their links, embedded with known elements (3). Here, we take into ac-count the person’s real-time brain state to optimize knowledge transfer: information is presented atmoments when theta oscillations indicate a favorable brain state for learning. We repeat this cycle (4)to identify non-encoded or forgotten information (yellow node, as opposed to the properly encodedorange nodes), until the brain representation matches the canonical knowledge representation.

A second potential domain for valorization to benefit society is education (Sigmanet al., 2014). Here, I propose a framework¹ for brain-based user models to enhance

¹Dubbed ”Full Crimea” for obvious reasons.

5

106

learning, and ultimately artificially engineer a person’s network of memories andknowledge base (Figure 5.2). The key method in this framework is similar to thememory network reconstruction approach showcased in Chapter 3. Although wewere unable to successfully utilize reconstructed memory networks on an individ-ual level, we provided essential groundwork for future research. Our contributionspave the way for potential improvements in individual memory network estimation.Assuming technological steps required to reliably read-out memory networks fromthe human brain have been made, we may employ the method to track knowledgeacquisition. Firstly, looking at the change of neural representations as a function oflearning may provide a measure of successful learning. Secondly, we may extractan individual’s neural knowledge network and compare it to a canonical knowledgestructure derived from large corpus of text, such as online encyclopedias. Akin toconcepts from personalized digital search, the neural knowledge structures informsa user model, which impacts search results or education material. Elements fromthe canonical knowledge structure that are missing in the neural knowledge struc-ture may be presented to the user. Maladaptive memories or misinformed neuralknowledge elements from a user may be disjoined from their neural network by em-phasizing other associations. This process may be used in conjunction with insightsand techniques from Chapter 2 and Chapter 4. The centrality of the hippocampus interms of network connectivity and amplitude of hippocampal theta oscillations maybe monitored to track brain states favorable for learning using a brain-computer in-terface (BCI). A high connector hub score for the hippocampus would indicate thatthe region is acting as a mnemonic convergence zone, whereas prominent thetaoscillations indicate that the hippocampus is ready to encode and retrieve memo-ries. More specifically, hippocampal theta phase, measured with MEG source recon-struction techniques may signal encoding and retrieval periods. Wemight speculateabout the possibility of entraining theta phase in certain brain regions, using min-imally invasive methods such as transcranial magnetic stimulation or transcranialalternating current stimulation, to achieve a favorable brain state for encoding ofnew information. Theta coupling with themPFCmight bemonitored to see whethernew information is integrated or rather segregated. All in all, the approach whereuser-specific missing knowledge is identified using representational analysis of neu-roimaging data in a BCI setting, may greatly accelerate learning by focusing learningefforts on missing pieces of information. In addition, identifying windows of favor-able encoding and integration brain states might further improve the efficacy of thelearning process.

5.4. Concluding remarksThe ability to form associations between a multitude of events is a hallmark ofepisodic memory. How does the brain perform this feat? Here, I outlined a se-

5

107

ries of experiments aimed to elucidate the neural mechanism of associative mem-ory, from several different methodological angles. The neuroscientific insights fromthese studies add to the knowledge on this topic, and the methods may aid futureresearch. In addition, these studies might have paved the way for innovative appli-cations in the clinical and educational domain.

Epilogue

November 7th, 2013, San Diego, California, USA.

Mental-time-travel back to the cruise. The hippocampal electrophysi-ologist concluded with some after-dinner science philosophy: ”What isthe most important question in science?” he asked, out of the blue. Thesomewhat surprised table party - including me - wittingly recited thekey research questions posited by pivotal figures in the field, includingthe master’s own. Without success. Finally, he decided to enlighten uswith his answer: ”The most important question, is your question”. How,in the name of Science, am I able to remember all of this?

109

References

Addante, R. J., Watrous, A. J., Yonelinas, A. P., Ekstrom, A. D., and Ranganath, C. (2011). Prestimulus thetaactivity predicts correct source memory retrieval. Proc. Natl. Acad. Sci. U S A., 108(26):10702–10707.

Andersen, P., Morris, R., Amaral, D., Bliss, T., and O’Keefe, J. (2006). The Hippocampus Book (OxfordNeuroscience Series). Oxford University Press.

Anderson, K. L., Rajagovindan, R., Ghacibeh, G. A., Meador, K. J., and Ding, M. (2010). Theta oscillationsmediate interaction between prefrontal cortex and medial temporal lobe in human memory. Cereb.Cortex, 20(7):1604–1612.

Attal, Y. and Schwartz, D. (2013). Assessment of subcortical source localization using deep brain activityimaging model with minimum norm operators: a MEG study. PLoS ONE, 8(3):e59856.

Avants, B. B., Tustison, N. J., Song, G., Cook, P. A., Klein, A., andGee, J. C. (2011). A reproducible evaluationof ANTs similarity metric performance in brain image registration. Neuroimage, 54(3):2033–2044.

Axmacher, N., Henseler, M. M., Jensen, O., Weinreich, I., Elger, C. E., and Fell, J. (2010). Cross-frequencycoupling supports multi-item working memory in the human hippocampus. Proc. Natl. Acad. Sci.U.S.A., 107(7):3228–3233.

Azab, M., Stark, S. M., and Stark, C. E. L. (2014). Contributions of human hippocampal subfields tospatial and temporal pattern separation. Hippocampus, 24:293–302.

Baillet, S., Mosher, J., and Leahy, R. (2001). Electromagnetic brain mapping. IEEE Signal Process. Mag.,18(6):14–30.

Barron, H. C., Dolan, R. J., and Behrens, T. E. J. (2013). Online evaluation of novel choices by simultaneousrepresentation of multiple memories. Nat. Neurosci., 16(10):1492–1498.

Batschelet, E. (1981). Circular Statistics in Biology (Mathematics in biology). Academic Press Inc.Battaglia, F. P., Benchenane, K., Sirota, A., Pennartz, C. M., and Wiener, S. I. (2011). The hippocampus:

hub of brain network communication for memory. Trends Cogn. Sci., 15:310–318.Benchenane, K., Peyrache, A., Khamassi, M., Tierney, P. L., Gioanni, Y., Battaglia, F. P., and Wiener, S. I.

(2010). Coherent theta oscillations and reorganization of spike timing in the hippocampal- pre-frontal network upon learning. Neuron, 66(6):921–936.

Berger, T. W., Alger, B., and Thompson, R. F. (1976). Neuronal substrate of classical conditioning in thehippocampus. Science, 192:483–485.

Binder, J. R. and Desai, R. H. (2011). The neurobiology of semantic memory. Trends Cogn. Sci., 15(11):527–536.

Biswal, B. B., Kylen, J. V., and Hyde, J. S. (1997). Simultaneous assessment of flow and BOLD signals inresting-state functional connectivity maps. NMR Biomed., 10(4-5):165–170.

Bliss, T. V. P. and Lømo, T. (1973). Long-lasting potentiation of synaptic transmission in the dentate areaof the anaesthetized rabbit following stimulation of the perforant path. J. Physiol., 232(2):331–356.

Blondel, V. D., Guillaume, J. L., Lambiotte, R., and Lefebvre, E. (2008). Fast unfolding of communities inlarge networks. J. Stat. Mech., 2008:P10008.

Bojangles, R. W. (2013). Sankrony: a novel single-trial phase coupling measure for imaginary electro-physiological signals. J. Cog. Neuroimaging.

Boring, E. G. (1950). A History of Experimental Psychology. Prentice Hall.

111

112

Bosch, S. E. (2016). Reactivating Memories in Hippocampus and Neocortex. PhD thesis, Radboud Uni-versity, Donders Graduate School for Cognitive Neuroscience.

Bosch, S. E., Jehee, J. F., Fernandez, G., and Doeller, C. F. (2014). Reinstatement of associative memoriesin early visual cortex is signaled by the hippocampus. J. Neurosci., 34(22):7493–7500.

Bota, M., Sporns, O., and Swanson, L. W. (2015). Architecture of the cerebral cortical association con-nectome underlying cognition. Proc. Natl Acad. Sci. USA, 112:E2093–E2101.

Brincat, S. L. and Miller, E. K. (2015). Frequency-specific hippocampal-prefrontal interactions duringassociative learning. Nat. Neurosci., 18(4):576–581.

Brodeur, M. B., Dionne-Dostie, E., Montreuil, T., and Lepage, M. (2010). The bank of standardized stimuli(BOSS), a new set of 480 normative photos of objects to be used as visual stimuli in cognitiveresearch. PLoS One, 5(5):e10773.

Buckner, R. L. (1996). Beyond HERA: Contributions of specific prefrontal brain areas to long-termmem-ory retrieval. Psychon. Bull. Rev., 3(2):149–158.

Buckner, R. L. and Krienen, F. M. (2013). The evolution of distributed association networks in the humanbrain. Trends Cogn. Sci., 17:648–665.

Buckner, R. L., Sepulcre, J., Talukdar, T., Krienen, F. M., Liu, H., Hedden, T., Andrews-Hanna, J. R., Sperling,R. A., and Johnson, K. A. (2009). Cortical hubs revealed by intrinsic functional connectivity: mapping,assessment of stability, and relation to Alzheimer’s disease. J. Neurosci., 29(6):1860–1873.

Bullmore, E. and Sporns, O. (2009). Complex brain networks: graph theoretical analysis of structuraland functional systems. Nat. Rev. Neurosci., 10:186–198.

Bunsey, M. and Elchenbaum, H. (1996). Conservation of hippocampal memory function in rats andhumans. Nature, 379(6562):255–257.

Burgess, N. (2002). The hippocampus, space, and viewpoints in episodic memory. Q. J. Exp. Psychol. A,55(4):1057–1080.

Buzsáki, G. (1996). The hippocampo-neocortical dialogue. Cerebral Cortex, 6(2):81–92.Buzsáki, G. (2002). Theta oscillations in the hippocampus. Neuron, 33(3):325–340.Buzsaki, G. (2005). Theta rhythmof navigation: link between path integration and landmark navigation,

episodic and semantic memory. Hippocampus, 15(7):827–840.Buzsaki, G. andMoser, E. I. (2013). Memory, navigation and theta rhythm in the hippocampal-entorhinal

system. Nat. Neurosci., 16(2):130–138.Cahusac, P. M. B., Rolls, E. T., Miyashita, Y., and Niki, H. (1993). Modification of the responses of

hippocampal neurons in the monkey during the learning of a conditional spatial response task. Hip-pocampus, 3(1):29–42.

Calkins, M. W. (1894). Experimental. Psych. Rev., 1(3):327–329.Canolty, R. T., Edwards, E., Dalal, S. S., Soltani, M., Nagarajan, S. S., Kirsch, H. E., Berger, M. S., Barbare,

N. M., and Knight, R. T. (2006). High gamma power is phase-locked to theta oscillations in humanneocortex. Science, 313(5793):1626–1628.

Capocchi, G., Zampolini, M., and Larson, J. (1992). Theta burst stimulation is optimal for induction of ltpat both apical and basal dendritic synapses on hippocampal ca1 neurons. Brain Res., 591(2):332–336.

Chadwick, M. J., Hassabis, D., Weiskopf, N., and Maguire, E. A. (2010). Decoding individual episodicmemory traces in the human hippocampus. Curr. Biol., 20:544–547.

Chen, J., Olsen, R. K., Preston, A. R., Glover, G. H., and Wagner, A. D. (2011). Associative retrievalprocesses in the human medial temporal lobe: Hippocampal retrieval success and CA1 mismatchdetection. Learn. Mem., 18(8):523–528.

Cohen, M. X. (2011). Hippocampal-prefrontal connectivity predicts midfrontal oscillations and long-term memory performance. Curr. Biol., 21(22):1900–1905.

Colgin, L. L., Moser, E. I., and Moser, M. B. (2008). Understanding memory through hippocampalremapping. Trends Neurosci., 31(9):469–477.

113

Collin, S. H., Milivojevic, B., and Doeller, C. F. (2015). Memory hierarchies map onto the hippocampallong axis in humans. Nat. Neurosci., 18:1562–1564.

Copara, M. S. (2014). Complementary roles of human hippocampal subregions during retrieval ofspatiotemporal context. J. Neurosci., 34:6834–6842.

Cornwell, B. R., Johnson, L. L., Holroyd, T., Carver, F. W., and Grillon, C. (2008). Human hippocampal andparahippocampal theta during goal-directed spatial navigation predicts performance on a virtualmorris water maze. J. Neurosci., 28(23):5983–5990.

Dalal, S. S., Jerbi, K., Bertrand, O., Adam, C., Ducorps, A., Schwartz, D., Martinerie, J., and Lachaux,J. P. (2013). Simultaneous meg-intracranial eeg: new insights into the ability of meg to captureoscillatory modulations in the neocortex and the hippocampus. Epilepsy Behav., 28:283–302.

Damasio, A. R. (1989). The brain binds entities and events bymultiregional activation from convergencezones. Neural Comput., 1:123–132.

Davachi, L. (2006). Item, context and relational episodic encoding in humans. Curr. Opin. Neurobiol.,16:693–700.

Davachi, L., Mitchell, J. P., and Wagner, A. D. (2003). Multiple routes to memory: Distinct medialtemporal lobe processes build item and source memories. Proc. Natl. Acad. Sci. U S A., 100(4):2157–2162.

de Gelder, B. and Van den Stock, J. (2011). The bodily expressive action stimulus test (beast). con-struction and validation of a stimulus basis for measuring perception of whole body expression ofemotions. Front. Psychol., 2:181.

de Kleijne, C. and Vito-Yoho, B. C. (1994). Compelling anecdotal evidence for neurocognitive superiorityof the first-born sibling. Acta Neuroapathetica.

Descartes, R. (1649). The Passions of the Soul: Les Passions De l’Âme. Hackett Publishing Company, Inc.DeVito, L. M., Lykken, C., Kanter, B. R., and Eichenbaum, H. (2010). Prefrontal cortex: Role in acquisition

of overlapping associations and transitive inference. Learn. Mem., 17(3):161–167.Dice, L. R. (1945). Measures of the amount of ecologic association between species. Ecology, 26:297–

302.Diedrichsen, J., Ridgway, G. R., Friston, K. J., and Wiestler, T. (2011). Comparing the similarity and spatial

structure of neural representations: A pattern-component model. Neuroimage, 55(4):1665–1678.Doeller, C. (2015). Cracking the code for memory: The doellerlab website.Doeller, C. F., Opitz, B., Krick, C. M., Mecklinger, A., and Reith, W. (2005). Prefrontal-hippocampal

dynamics involved in learning regularities across episodes. Cereb. Cortex, 15(8):1123–1133.Dragoi, G., Carpi, D., Recce, M., Csicsvari, J., and Buzsaki, G. (1999). Interactions between hippocam-

pus and medial septum during sharp waves and theta oscillation in the behaving rat. J. Neurosci.,19(14):6191–6199.

Durstewitz, D., Vittoz, N. M., Floresco, S. B., and Seamans, J. K. (2010). Abrupt transitions betweenprefrontal neural ensemble states accompany behavioral transitions during rule learning. Neuron,66(3):438–448.

Dusek, J. A. and Eichenbaum, H. (1997). The hippocampus and memory for orderly stimulus relations.Proc. Natl. Acad. Sci. U S A., 94(13):7109–7114.

Eichenbaum, H. (2000). A cortical-hippocampal system for declarative memory. Nat. Rev. Neurosci.,1:41–50.

Eichenbaum, H., Dudchenko, P., Wood, E., Shapiro, M., and Tanila, H. (1999). The hippocampus, memory,and place cells: Is it spatial memory or a memory space? Neuron, 23(2):209–226.

Eichenbaum, H., Wiener, S. I., Shapiro, M. L., and Cohen, N. J. (1989). The organization of spatial codingin the hippocampus: a study of neural ensemble activity. J. Neurosci., 9:2764–2775.

Eichenbaum, H., Yonelinas, A. P., and Ranganath, C. (2007). The medial temporal lobe and recognitionmemory. Annu. Rev. Neurosci., 30:123–152.

114

Ekman, M., Derrfuss, J., Tittgemeyer, M., and Fiebach, C. J. (2012). Predicting errors from reconfigurationpatterns in human brain networks. Proc. Natl Acad. Sci. USA, 109:16714–16719.

Eldridge, L. L., Knowlton, B. J., Furmanski, C. S., Bookheimer, S. Y., and Engel, S. A. (2000). Rememberingepisodes: a selective role for the hippocampus during retrieval. Nat. Neurosci., 3(11):1149–1152.

Erickson, C. A., Jagadeesh, B., and Desimone, R. (2000). Clustering of perirhinal neurons with similarproperties following visual experience in adult monkeys. Nat. Neurosci., 3(11):1143–1148.

Euston, D. R., Gruber, A. J., and McNaughton, B. L. (2012). The role of medial prefrontal cortex inmemory and decision making. Neuron, 76(6):1057–1070.

Fell, J. and Axmacher, N. (2011). The role of phase synchronization in memory processes. Nat. Rev.Neurosci., 12(2):105–118.

Felleman, D. J. and Van Essen, D. C. (1991). Distributed hierarchical processing in the primate cerebralcortex. Cereb. Cortex, 1:1–47.

Fi, L. A. I. and Béziers, M. F. (2014). Where are daddy’s ears? A remarkable case of transitive inferencein an 17-month old human infant. Psych. Enfant.

Filler, A. (2009). MR neurography and diffusion tensor imaging: Origins, history & clinical impact. Nat.Prec.

Fletcher, P., Frith, C., Baker, S., Shallice, T., Frackowiak, R., and Dolan, R. (1995). The mind’s eye—precuneus activation in memory-related imagery. Neuroimage, 2(3):195–200.

Frankland, P. W. and Bontempi, B. (2005). The organization of recent and remote memories. Nat. Rev.Neurosci., 6(2):119–130.

Freeman, J. (2015). Open source tools for large-scale neuroscience. Curr. Opin. Neurobiol., 32:156–163.Fries, P. (2005). A mechanism for cognitive dynamics: neuronal communication through neuronal

coherence. Trends Cogn. Sci., 9(10):474–480.Friston, K. J. (2011). Functional and effective connectivity: A review. Brain Connect., 1(1):13–36.Fuentemilla, L., Barnes, G. R., Duzel, E., and Levine, B. (2014). Theta oscillations orchestrate medial

temporal lobe and neocortex in remembering autobiographicalmemories. Neuroimage, 85 Pt 2:730–737.

Fujisawa, S. and Buzáski, G. (2011). A 4 hz oscillation adaptively synchronizes prefrontal, VTA, andhippocampal activities. Neuron, 72(1):153–165.

Garrido, M. I., Barnes, G. R., Kumaran, D., Maguire, E. A., and Dolan, R. J. (2015). Ventromedial pre-frontal cortex drives hippocampal theta oscillations induced by mismatch computations. Neuroim-age, 120:362–370.

Grastyan, E., Lissak, K., Madarasz, I., and Donhoffer, H. (1959). Hippocampal electrical activity duringthe development of conditioned reflexes. Electroencephalogr. Clin. Neurophysiol., 11(3):409–430.

Greene, A. J., Gross, W. L., Elsinger, C. L., and Rao, S. M. (2006). An fMRI analysis of the human hip-pocampus: Inference, context, and task awareness. J. Cogn. Neurosci., 18(7):1156–1173.

Gross, J., Kujala, J., Hämäläinen, M., Timmermann, L., Schnitzler, A., and Salmelin, R. (2001). Dynamicimaging of coherent sources: Studying neural interactions in the human brain. Proc. Natl. Acad. Sci.U S A., 98(2):694–699.

Guderian, S., Schott, B. H., Richardson-Klavehn, A., and Duzel, E. (2009). Medial temporal thetastate before an event predicts episodic encoding success in humans. Proc. Natl. Acad. Sci. U.S.A.,106(13):5365–5370.

Guimera, R., Mossa, S., Turtschi, A., and Amaral, L. A. (2005). The worldwide air transportation net-work: anomalous centrality, community structure, and cities’ global roles. Proc. Natl Acad. Sci. USA,102:7794–7799.

Guitart-Masip, M., Barnes, G. R., Horner, A., Bauer, M., Dolan, R. J., and Duzel, E. (2013). Synchronizationof medial temporal lobe and prefrontal rhythms in human decision making. J. Neurosci., 33(2):442–451.

115

Gutchess, A. H. and Schacter, D. L. (2012). The neural correlates of gist-based true and false recognition.Neuroimage, 59:3418–3426.

Hannula, D. E. and Ranganath, C. (2008). Medial temporal lobe activity predicts successful relationalmemory binding. J. Neurosci., 28:116–124.

Hanslmayr, S. and Staudigl, T. (2014). How brain oscillations form memories - a processing basedperspective on oscillatory subsequent memory effects. Neuroimage, 85:648–655.

Hanslmayr, S., Staudigl, T., and Fellner, M.-C. (2012). Oscillatory power decreases and long-term mem-ory: the information via desynchronization hypothesis. Front. Hum. Neurosci., 6.

Hasselmo, M. E., Bodelón, C., and Wyble, B. P. (2002). A proposed function for hippocampal thetarhythm: Separate phases of encoding and retrieval enhance reversal of prior learning. Neural Com-put., 14(4):793–817.

Haynes, J.-D. (2015). A primer on pattern-based approaches to fMRI: Principles, pitfalls, and perspec-tives. Neuron, 87(2):257–270.

Haynes, J.-D. and Rees, G. (2006). Decoding mental states from brain activity in humans. Nat. Rev.Neurosci., 7(7):523–534.

Hebb, D. (1949). The Organization of Behavior. New York: Wiley.Heckers, S., Zalesak, M., Weiss, A. P., Ditman, T., and Titone, D. (2004). Hippocampal activation during

transitive inference in humans. Hippocampus, 14(2):153–162.Henson, R. N. A., Shallice, T., and Dolan, R. J. (1999). Right prefrontal cortex and episodic memory

retrieval: a functional MRI test of the monitoring hypothesis. Brain, 122(7):1367–1381.Hinne, M., Janssen, R. J., Heskes, T., and van Gerven, M. A. (2015). Bayesian Estimation of Con-

ditional Independence Graphs Improves Functional Connectivity Estimates. PLoS Comput. Biol.,11(11):e1004534.

Hipp, J. F., Hawellek, D. J., Corbetta,M., Siegel, M., and Engel, A. K. (2012). Large-scale cortical correlationstructure of spontaneous oscillatory activity. Nat. Neurosci., 15(6):884–890.

Hoesen, G. W. V., Pandya, D. N., and Butters, N. (1972). Cortical afferents to the entorhinal cortex ofthe rhesus monkey. Science, 175(4029):1471–1473.

Horner, A. J., Bisby, J. A., Bush, D., Lin, W. J., and Burgess, N. (2015). Evidence for holistic episodicrecollection via hippocampal pattern completion. Nat. Commun., 6:7462.

Horner, A. J. and Burgess, N. (2014). Pattern completion in multielement event engrams. Curr. Biol.,24:988–992.

Hsieh, L.-T., Gruber, M. J., Jenkins, L. J., and Ranganath, C. (2014). Hippocampal activity patterns carryinformation about objects in temporal context. Neuron, 81(5):1165–1178.

Huettel, S. A., Song, A. W., and McCarthy, G. (2008). Functional Magnetic Resonance Imaging, SecondEdition. Sinauer Associates.

Hulse, S. and Deese, J. (1967). The Psychology of Learning. McGraw-Hill.Hyman, J. M., Ma, L., Balaguer-Ballester, E., Durstewitz, D., and Seamans, J. K. (2012). Contextual encod-

ing by ensembles of medial prefrontal cortex neurons. Proc. Natl. Acad. Sci. U.S.A., 109(13):5086–5091.

Hyman, J.M., Wyble, B. P., Goyal, V., Rossi, C. A., andHasselmo,M. E. (2003). Stimulation in hippocampalregion CA1 in behaving rats yields long-term potentiation when delivered to the peak of theta andlong-term depression when delivered to the trough. J. Neurosci., 23(37):11725–11731.

Hyman, J. M., Zilli, E. A., Paley, A. M., and Hasselmo, M. E. (2005). Medial prefrontal cortex cells showdynamic modulation with the hippocampal theta rhythm dependent on behavior. Hippocampus,15(6):739–749.

Ison, M. J., Quiroga, R. Q., and Fried, I. (2015). Rapid encoding of new memories by individual neuronsin the human brain. Neuron, 87(1):220–230.

Jacobs, J. (2014). Hippocampal theta oscillations are slower in humans than in rodents: Implicationsfor models of spatial navigation and memory. Philos. Trans. R Soc. Lond. B Biol. Sci., 369(1635).

116

Jay, T. M. and Witter, M. P. (1991). Distribution of hippocampal ca1 and subicular efferents in theprefrontal cortex of the rat studied by means of anterograde transport of phaseolus vulgaris-leucoagglutinin. J. Comp. Neurol., 313(4):574–586.

Jenkinson, M., Bannister, P., Brady, M., and Smith, S. (2002). Improved optimization for the robust andaccurate linear registration and motion correction of brain images. Neuroimage, 17(2):825–841.

Jezek, K., Henriksen, E. J., Treves, A., Moser, E. I., and Moser, M. . (2011). Theta-paced flickering betweenplace-cell maps in the hippocampus. Nature, 478(7368):246–249.

Jo, Y. S., Park, E. H., Kim, I. H., Park, S. K., Kim, H., Kim, H. T., and Choi, J.-S. (2007). The medial prefrontalcortex is involved in spatial memory retrieval under partial-cue conditions. J. Neurosci., 27(49):13567–13578.

Jones, M. W. and Wilson, M. A. (2005). Phase precession of medial prefrontal cortical activity relativeto the hippocampal theta rhythm. Hippocampus, 15(7):867–873.

Jouvet, M. (1969). Biogenic amines and the states of sleep. Science, 163(3862):32–41.Kamitani, Y. and Tong, F. (2005). Decoding the visual and subjective contents of the human brain. Nat.

Neurosci., 8(5):679–685.Kaplan, R., Bush, D., Bonnefond, M., Bandettini, P. A., Barnes, G. R., Doeller, C. F., and Burgess, N. (2014).

Medial prefrontal theta phase coupling during spatial memory retrieval. Hippocampus, 24(6):656–665.

Klemm, W. R. (1972). Effects of electric stimulation of brain stem reticular formation on hippocampaltheta rhythm and muscle activity in unanesthetized, cervical- and midbrain-transected rats. BrainRes., 41(2):331–344.

Komorowski, R. W., Manns, J. R., and Eichenbaum, H. (2009). Robust conjunctive item-place coding byhippocampal neurons parallels learning what happens where. J. Neurosci., 29(31):9918–9929.

Koscik, T. R. and Tranel, D. (2012). The human ventromedial prefrontal cortex is critical for transitiveinference. J. Cogn. Neurosci., 24(5):1191–1204.

Kriegeskorte, N., Goebel, R., and Bandettini, P. (2006). Information-based functional brain mapping.Proc. Natl Acad. Sci. USA, 103:3863–3868.

Kriegeskorte, N. and Mur, M. (2012). Inverse MDS: Inferring dissimilarity structure from multiple itemarrangements. Front. Psychol., 3.

Kuhl, B. A., Johnson, M. K., and Chun, M. M. (2013). Dissociable neural mechanisms for goal-directedversus incidental memory reactivation. Jo. Neurosci., 33(41):16099–16109.

Kumaran, D. and McClelland, J. L. (2012). Generalization through the recurrent interaction of episodicmemories: A model of the hippocampal system. Psychol. Rev, 119(3):573–616.

Kumaran, D., Summerfield, J. J., Hassabis, D., and Maguire, E. A. (2009). Tracking the emergence ofconceptual knowledge during human decision making. Neuron, 63(6):889–901.

Lachaux, J.-P., Rodriguez, E., Martinerie, J., and Varela, F. J. (1999). Measuring phase synchrony in brainsignals. Hum. Brain Mapp., 8(4):194–208.

LaRocque, K. F. (2013). Global similarity and pattern separation in the human medial temporal lobepredict subsequent memory. J. Neurosci., 33:5466–5474.

Lashley, K. (1950). In search of the engram. Symp. Soc. Exp. Biol., 4:454–482.Lavenex, P. and Amaral, D. G. (2000). Hippocampal-neocortical interaction: a hierarchy of associativity.

Hippocampus, 10:420–430.Leemput, K. V., Bakkour, A., Benner, T., Wiggins, G., Wald, L. L., Augustinack, J., Dickerson, B. C., Gol-

land, P., and Fischl, B. (2009). Automated segmentation of hippocampal subfields from ultra-highresolution in vivo MRI. Hippocampus, 19(6):549–557.

Lega, B. C., Jacobs, J., and Kahana, M. (2012). Human hippocampal theta oscillations and the formationof episodic memories. Hippocampus, 22(4):748–761.

Lewis, C. M., Baldassarre, A., Committeri, G., Romani, G. L., and Corbetta, M. (2009). Learning sculptsthe spontaneous activity of the resting human brain. Proc. Natl Acad. Sci. USA, 106:17558–17563.

117

Limpiti, T., Van Veen, B. D., and Wakai, R. T. (2006). Cortical patch basis model for spatially extendedneural activity. IEEE Trans. Biomed. Eng., 53(9):1740–1754.

Lo, A., Chernoff, H., Zheng, T., and Lo, S.-H. (2015). Why significant variables aren’t automatically goodpredictors. Proc. Natl. Acad. Sci. U S A., 112(45):13892–13897.

Logothetis, N. K., Pauls, J., Augath, M., Trinath, T., and Oeltermann, A. (2001). Neurophysiologicalinvestigation of the basis of the fmri signal. Nature, 412(6843):150–157.

Long, N. M., Burke, J. F., and Kahana, M. J. (2014). Subsequent memory effect in intracranial and scalpeeg. Neuroimage, 84:488–494.

Lundqvist, D., Flykt, A., and Ohman, A. (1998). The Karolinska Directed Emotional Faces (KDEF).Manjon, J. V., Coupe, P., Marti-Bonmati, L., Collins, D. L., and Robles, M. (2010). Adaptive non-local

means denoising of mr images with spatially varying noise levels. J. Magn. Reson. Imaging, 31:192–203.

Manns, J. R. and Eichenbaum, H. (2009). A cognitive map for object memory in the hippocampus.Learn. Mem., 16(10):616–624.

Manns, J. R., Howard, M. W., and Eichenbaum, H. (2007). Gradual changes in hippocampal activitysupport remembering the order of events. Neuron, 56(3):530–540.

Maris, E., Schoffelen, J.-M., and Fries, P. (2007). Nonparametric statistical testing of coherence differ-ences. J. Neurosci. Methods., 163(1):161–175.

Marr, D. (1971). Simple memory: a theory for archicortex. Philos. Trans. R. Soc. Lond. B Biol. Sci., 262:23–81.

McClelland, J. L. (1994). The organization of memory. a parallel distributed processing perspective. Rev.Neurol. (Paris), 150:570–579.

McKenzie, S., Frank, A. J., Kinsky, N. R., Porter, B., Rivière, P. D., and Eichenbaum, H. (2014). Hippocampalrepresentation of related and opposing memories develop within distinct, hierarchically organizedneural schemas. Neuron, 83(1):202–215.

McKenzie, S., Robinson, N. T. M., Herrera, L., Churchill, J. C., and Eichenbaum, H. (2013). Learning causesreorganization of neuronal firing patterns to represent related experiences within a hippocampalschema. J. Neurosci., 33(25):10243–10256.

McNaughton, B. and Morris, R. (1987). Hippocampal synaptic enhancement and information storagewithin a distributed memory system. Trends Neurosci., 10(10):408–415.

Mielke, M. M., Kozauer, N. A., Chan, K. C., George, M., Toroney, J., Zerrate, M., Bandeen-Roche, K.,Wang, M. C., Vanzijl, P., Pekar, J. J., Mori, S., Lyketsos, C. G., and Albert, M. (2009). Regionally-specific diffusion tensor imaging inmild cognitive impairment andAlzheimer’s disease.Neuroimage,46(1):47–55.

Milivojevic, B. and Doeller, C. F. (2013). Mnemonic networks in the hippocampal formation: Fromspatial maps to temporal and conceptual codes. J. Exp. Psychol. Gen., 142(4):1231–1241.

Milivojevic, B., Vicente-Grabovetsky, A., and Doeller, C. F. (2015). Insight reconfigures hippocampal-prefrontal memories. Curr. Biol., 25:821–830.

Miller, E. K. (2000). The prefontral cortex and cognitive control. Nat. Rev. Neurosci., 1(1):59–65.Misic, B., Goni, J., Betzel, R. F., Sporns, O., and McIntosh, A. R. (2014). A network convergence zone in

the hippocampus. PLoS Comput. Biol., 10:e1003982.Moita, M. A., Rosis, S., Zhou, Y., LeDoux, J. E., and Blair, H. T. (2003). Hippocampal place cells acquire

location-specific responses to the conditioned stimulus during auditory fear conditioning. Neuron,37:485–497.

Monaco, J. D., Knierim, J. J., and Zhang, K. (2011). Sensory feedback, error correction, and remapping ina multiple oscillator model of place-cell activity. Front. Comput. Neurosci., (SEPTEMBER).

Moser, M.-B. and Moser, E. I. (2016). The brain’s gps tells you where you are and where you’ve comefrom. Scientific American, January.

118

Mumford, J. A., Davis, T., and Poldrack, R. A. (2014). The impact of study design on pattern estimationfor single-trial multivariate pattern analysis. Neuroimage, 103:130–138.

Mumford, J. A., Turner, B. O., Ashby, F. G., and Poldrack, R. A. (2012). Deconvolving bold activation inevent-related designs for multivoxel pattern classification analyses. Neuroimage, 59:2636–2643.

Murre, J. M. (1996). TraceLink: A model of amnesia and consolidation of memory. Hippocampus,6(6):675–684.

Nadel, L. and Peterson, M. A. (2013). The hippocampus: part of an interactive posterior representationalsystem spanning perceptual and memorial systems. J. Exp. Psychol. Gen., 142:1242–1254.

Nagode, J. C. and Pardo, J. V. (2002). Human hippocampal activation during transitive inference. Neu-roReport, 13(7):939–944.

Navarro-Schröder, T. (2016). Grids and Gradients. PhD thesis, Radboud University, Donders GraduateSchool for Cognitive Neuroscience.

Navarro-Schroder, T., Haak, K. V., Jimenez, N. I. Z., Beckmann, C. F., and Doeller, C. F. (2015). Functionaltopography of the human entorhinal cortex. Elife, 4:e06738.

Nolde, S. F., Johnson, M. K., and Raye, C. L. (1998). The role of prefrontal cortex during tests of episodicmemory. Trends Cogn. Sci., 2(10):399–406.

Nolte, G. (2003). The magnetic lead field theorem in the quasi-static approximation and its use formagnetoenchephalography forward calculation in realistic volume conductors. Phys. Med. Biol.,48(22):3637–3652.

Oehrn, C. R., Baumann, C., Fell, J., Lee, H., Kessler, H., Habel, U., Hanslmayr, S., and Axmacher, N. (2015).Human hippocampal dynamics during response conflict. Curr. Biol., 25(17):2307–2313.

O’Keefe, J. and Dostrovsky, J. (1971). The hippocampus as a spatial map. preliminary evidence from unitactivity in the freely-moving rat. Brain Res., 34(1):171–175.

O’Keefe, J. and Nadel, L. (1978). The Hippocampus as a Cognitive Map.O’Keefe, J. and Recce, M. L. (1993). Phase relationship between hippocampal place units and the eeg

theta rhythm. Hippocampus, 3(3):317–330.Oliva, A. and Torralba, A. (2001). Modeling the shape of the scene: A holistic representation of the

spatial envelope. Int. J. Comput. Vis., 42:145–175.Öngür, D., Ferry, A. T., and Price, J. L. (2003). Architectonic subdivision of the human orbital and medial

prefrontal cortex. J. Comp. Neurol., 460(3):425–449.Oostenveld, R., Fries, P., Maris, E., and Schoffelen, J.-M. (2011). FieldTrip: Open source software for

advanced analysis of MEG, EEG, and invasive electrophysiological data. Comput. Intell. Neurosci.,2011:1–9.

O’Reilly, R. C. and Rudy, J. W. (2001). Conjunctive representations in learning and memory: principlesof cortical and hippocampal function. Psychol. Rev., 108:311–345.

Paller, K. A. andWagner, A. D. (2002). Observing the transformation of experience intomemory. TrendsCogn. Sci., 6(2):93–102.

Papez, J. W. (1944). Structures and mechanisms underlying the cerebral functions. Am. J. Psychol.,57:291–316.

Pastalkova, E., Itskov, V., Amarasingham, A., and Buzsaki, G. (2008). Internally generated cell assemblysequences in the rat hippocampus. Science, 321(5894):1322–1327.

Patterson, K., Nestor, P. J., and Rogers, T. T. (2007). Where do you know what you know? the represen-tation of semantic knowledge in the human brain. Nat. Rev. Neurosci., 8(12):976–987.

Peelen, M. V. and Kastner, S. (2014). Attention in the real world: toward understanding its neural basis.Trends Cogn. Sci., 18(5):242–250.

Poppenk, J., Evensmoen, H. R., Moscovitch, M., and Nadel, L. (2013). Long-axis specialization of thehuman hippocampus. Trends Cogn. Sci., 17:230–240.

Poser, B. A., Koopmans, P. J., Witzel, T., Wald, L. L., and Barth, M. (2010). Three dimensional echo-planarimaging at 7 tesla. Neuroimage, 51:261–266.

119

Power, J. D., Schlaggar, B. L., Lessov-Schlaggar, C. N., and Petersen, S. E. (2013). Evidence for hubs inhuman functional brain networks. Neuron, 79:798–813.

Preston, A. R. and Eichenbaum, H. (2013). Interplay of hippocampus and prefrontal cortex in memory.Curr. Biol., 23(17):R764–R773.

Preston, A. R., Shrager, Y., Dudukovic, N. M., and Gabrieli, J. D. (2004). Hippocampal contribution tothe novel use of relational information in declarative memory. Hippocampus, 14(2):148–152.

Quiroga, R. Q., Reddy, L., Kreiman, G., Koch, C., and Fried, I. (2005). Invariant visual representation bysingle neurons in the human brain. Nature, 435:1102–1107.

Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, D. A., and Shulman, G. L. (2001). Adefault mode of brain function. Proc. Natl. Acad. Sci. U S A., 98(2):676–682.

Rajasethupathy, P., Sankaran, S., Marshel, J. H., Kim, C. K., Ferenczi, E., Lee, S. Y., Berndt, A., Ramakrish-nan, C., Jaffe, A., Lo, M., Liston, C., and Deisseroth, K. (2015). Projections from neocortex mediatetop-down control of memory retrieval. Nature, 526(7575):653–659.

Ranganath, C., Cohen, M. X., Dam, C., and D’Esposito, M. (2004). Inferior temporal, prefrontal, and hip-pocampal contributions to visual working memory maintenance and associative memory retrieval.J. Neurosci., 24:3917–3925.

Ranganath, C. and Paller, K. A. (1999). Frontal brain potentials during recognition are modulated byrequirements to retrieve perceptual detail. Neuron, 22(3):605–613.

Ranganath, C. and Ritchey, M. (2012). Two cortical systems for memory-guided behaviour. Nat. Rev.Neurosci., 13:713–726.

Reber, P. (2010). What is the memory capacity of the human brain? Scientific American, May.Reddy, L., Poncet, M., Self, M. W., Peters, J. C., Douw, L., van Dellen, E., Claus, S., Reijneveld, J. C., Baayen,

J. C., and Roelfsema, P. R. (2015). Learning of anticipatory responses in single neurons of the humanmedial temporal lobe. Nat. Commun., 6:8556.

Riggs, L., Moses, S. N., Bardouille, T., Herdman, A. T., Ross, B., and Ryan, J. D. (2009). A complemen-tary analytic approach to examining medial temporal lobe sources using magnetoencephalography.Neuroimage, 45(2):627–642.

Rissman, J., Gazzaley, A., and D’Esposito, M. (2004). Measuring functional connectivity during distinctstages of a cognitive task. Neuroimage, 23:752–763.

Rissman, J. and Wagner, A. D. (2012). Distributed representations in memory: Insights from functionalbrain imaging. Annu. Rev. Psychol., 63(1):101–128.

Ritchey, M., Yonelinas, A. P., and Ranganath, C. (2014). Functional connectivity relationships predictsimilarities in task activation and pattern information during associative memory encoding. J. Cogn.Neurosci., 26:1085–1099.

Roediger, H. L., Dudai, Y., and Fitzpatrick, S. M. (2007). Science of Memory Concepts. Oxford UniversityPress.

Rugg, M. D., Fletcher, P. C., Chua, P. M.-L., and Dolan, R. J. (1999). The role of the prefrontal cortex inrecognition memory and memory for source: An fMRI study. Neuroimage, 10(5):520–529.

Rutishauser, U., Ross, I. B., Mamelak, A. N., and Schuman, E. M. (2010). Human memory strength ispredicted by theta-frequency phase-locking of single neurons. Nature, 464(7290):903–907.

Sakai, K. and Miyashita, Y. (1991). Neural organization for the long-term memory of paired associates.Nature, 354(6349):152–155.

Sargolini, F. (2006). Conjunctive representation of position, direction, and velocity in entorhinal cortex.Science, 312:758–762.

Schacter, D. L. (2001). Forgotten Ideas, Neglected Pioneers: Richard Semon and the Story of Memory.Psychology Press.

Schapiro, A. C., Kustner, L. V., and Turk-Browne, N. B. (2012). Shaping of object representations in thehuman medial temporal lobe based on temporal regularities. Curr. Biol., 22(17):1622–1627.

120

Schapiro, A. C., Rogers, T. T., Cordova, N. I., Turk-Browne, N. B., and Botvinick, M. M. (2013). Neuralrepresentations of events arise from temporal community structure. Nat. Neurosci., 16(4):486–492.

Schapiro, A. C., Turk-Browne, N. B., Norman, K. A., and Botvinick, M. M. (2015). Statistical learning oftemporal community structure in the hippocampus. Hippocampus, 26(1):3–8.

Schedlbauer, A. M., Copara, M. S., Watrous, A. J., and Ekstrom, A. D. (2014). Multiple interacting brainareas underlie successful spatiotemporal memory retrieval in humans. Sci. Rep., 4:6431.

Schlegel, A. (2013). Network structure and dynamics of the mental workspace. Proc. Natl Acad. Sci.USA, 110:16277–16282.

Schlichting, M. L., Mumford, J. A., and Preston, A. R. (2015). Learning-related representational changesreveal dissociable integration and separation signatures in the hippocampus and prefrontal cortex.Nat. Commun., 6:8151.

Scoville, W. B. andMilner, B. (1957). Loss of recentmemory after bilateral hippocampal lesions. J. Neurol.Neurosurg. Psychiatry, 20(1):11–21.

Sederberg, P. B., Kahana, M. J., Howard, M. W., Donner, E. J., and Madsen, J. R. (2003). Theta and gammaoscillations during encoding predict subsequent recall. J. Neurosci., 23(34):10809–10814.

Semon, R. W. (1921). The nmeme. London: George Allen and Unwin.Semon, R. W. (1923). The mneme. London: George Allen and Unwin.Shohamy, D. and Turk-Browne, N. B. (2013). Mechanisms for widespread hippocampal involvement in

cognition. J. Exp. Psychol. Gen., 142:1159–1170.Shohamy, D. and Wagner, A. D. (2008). Integrating memories in the human brain: Hippocampal-

midbrain encoding of overlapping events. Neuron, 60(2):378–389.Siapas, A. G., Lubenov, E. V., and Wilson, M. A. (2005). Prefrontal phase locking to hippocampal theta

oscillations. Neuron, 46(1):141–151.Siegel, M., Donner, T. H., and Engel, A. K. (2012). Spectral fingerprints of large-scale neuronal interac-

tions. Nat. Rev. Neurosci., 13(2):121–134.Sigman, M., Pena, M., Goldin, A. P., and Ribeiro, S. (2014). Neuroscience and education: prime time to

build the bridge. Nat. Neurosci., 17(4):497–502.Sigurdsson, T., Stark, K. L., Karayiorgou, M., Gogos, J. A., andGordon, J. A. (2010). Impaired hippocampal-

prefrontal synchrony in a genetic mouse model of schizophrenia. Nature, 464(7289):763–767.Simian-Orange, J., Dock, P. P., Wabb, A. S. C., and Del Erizo, Z. (2011). The detrimental effects of Horse

Race Betting and deceit-based Small Trades on hippocampal function: a monkey fMRI study. Ann.Sci. Mon. Ult. Wee.

Simons, J. S. and Spiers, H. J. (2003). Prefrontal and medial temporal lobe interactions in long-termmemory. Nat. Rev. Neurosci., 4(8):637–648.

Sirota, A., Montgomery, S., Fujisawa, S., Isomura, Y., Zugaro, M., and Buzsaki, G. (2008). Entrainment ofneocortical neurons and gamma oscillations by the hippocampal theta rhythm. Neuron, 60(4):683–697.

Smith, S. M. (2002). Fast robust automated brain extraction. Hum. Brain Mapp., 17(3):143–155.Smith, S. M. and Nichols, T. E. (2009). Threshold-free cluster enhancement: addressing problems of

smoothing, threshold dependence and localisation in cluster inference. Neuroimage, 44:83–98.Squire, L. R. (1987). Memory and Brain. Oxford University Press.Standing, L., Conezio, J., and Haber, R. N. (1970). Perception and memory for pictures: Single-trial

learning of 2500 visual stimuli. Psychon. Sci., 19(2):73–74.Staresina, B. P., Cooper, E., and Henson, R. N. (2013). Reversible information flow across the me-

dial temporal lobe: The hippocampus links cortical modules during memory retrieval. J. Neurosci.,33(35):14184–14192.

Staresina, B. P., Henson, R. N., Kriegeskorte, N., and Alink, A. (2012). Episodic reinstatement in themedial temporal lobe. J. Neurosci., 32:18150–18156.

121

Stark, C. E. (2002). Recognition memory for single items and for associations is similarly impairedfollowing damage to the hippocampal region. Learn. Mem., 9(5):238–242.

Stark, C. E. and Squire, L. R. (2001). Simple and associative recognition memory in the hippocampalregion. Learn. Mem., 8:190–197.

Staudigl, T. and Hanslmayr, S. (2013). Theta oscillations at encoding mediate the context-dependentnature of human episodic memory. Curr. Biol., 23(12):1101–1106.

Steemers, B., Vicente-Grabovetsky, A., Barry, C., Smulders, P., Schröder, T. N., Burgess, N., and Doeller,C. F. (2016). Hippocampal attractor dynamics predict memory-based decision making. Curr. Biol.

Stephen, J. M., Ranken, D. M., Aine, C. J., Weisend, M. P., and Shih, J. J. (2005). Differentiability ofsimulated MEG hippocampal, medial temporal and neocortical temporal epileptic spike activity. J.Clin. Neurophysiol., 22(6):388–401.

Stewart, M. and Fox, S. E. (1990). Firing relations of lateral septal neurons to the hippocampal thetarhythm in urethane anesthetized rats. Exp. Brain Res., 79(1):92–96.

Stokes, M. G., Kusunoki, M., Sigala, N., Nili, H., Gaffan, D., and Duncan, J. (2013). Dynamic coding forcognitive control in prefrontal cortex. Neuron, 78(2):364–375.

Stokes, M. G., Wolff, M. J., and Spaak, E. (2015). Decoding Rich Spatial Information with High TemporalResolution. Trends Cogn. Sci., 19(11):636–638.

Stolk, A., Todorovic, A., Schoffelen, J.-M., and Oostenveld, R. (2013). Online and offline tools for headmovement compensation in MEG. Neuroimage, 68:39–48.

Takashima, A., Petersson, K. M., Rutters, F., Tendolkar, I., Jensen, O., Zwarts, M. J., McNaughton, B. L.,and Fernandez, G. (2006). Declarative memory consolidation in humans: a prospective functionalmagnetic resonance imaging study. Proc. Natl. Acad. Sci. U.S.A., 103(3):756–761.

Tenenbaum, J. B., Kemp, C., Griffiths, T. L., and Goodman, N. D. (2011). How to grow a mind: Statistics,structure, and abstraction. Science, 331(6022):1279–1285.

Teyler, T. J. and DiScenna, P. (1986). The hippocampal memory indexing theory. Behav. Neurosci.,100:147–154.

Tolman, E. C. (1948). Cognitive maps in rats and men. Psychol. Rev., 55(4):189–208.Tonegawa, S., Liu, X., Ramirez, S., and Redondo, R. (2015). Memory Engram Cells Have Come of Age.

Neuron, 87(5):918–931.Treves, A. and Rolls, E. T. (1994). Computational analysis of the role of the hippocampus in memory.

Hippocampus, 4(3):374–391.Tse, D., Langston, R. F., Kakeyama, M., Bethus, I., Spooner, P. A., Wood, E. R., Witter, M. P., and Morris, R.

G. M. (2007). Schemas and memory consolidation. Science, 316(5821):76–82.Tulving, E., Donaldson, W., Bower, G., and of Naval Research, U. S. O. (1972). Organization of memory.

Academic Press.Tulving, E., Voi, M. E. L., Routh, D. A., and Loftus, E. (1983). Ecphoric processes in episodic memory.

Philos. Trans. R. Soc. Lond. B Biol. Sci., 302(1110):361–371.Tulving, E. and Watkins, M. J. (1975). Structure of memory traces. Psychol. Rev., 82:261–275.Turgut, Y. B. and Turgut, M. (2011). A mysterious term hippocampus involved in learning and memory.

Childs Nerv. Syst., 27(12):2023–2025.Tzourio-Mazoyer, N. (2002). Automated anatomical labeling of activations in spm using a macroscopic

anatomical parcellation of the mni mri single-subject brain. Neuroimage, 15:273–289.van de Nieuwenhuijzen, M. E., Backus, A. R., Bahramisharif, A., Doeller, C. F., Jensen, O., and van Gerven,

M. A. (2013). MEG-based decoding of the spatiotemporal dynamics of visual category perception.Neuroimage, 83:1063–1073.

van den Heuvel, M. P. and Sporns, O. (2011). Rich-club organization of the human connectome. J.Neurosci., 31:15775–15786.

van den Heuvel, M. P. and Sporns, O. (2013). Network hubs in the human brain. Trends Cogn. Sci.,17(12):683–696.

122

van Gerven, M., Farquhar, J., Schaefer, R., Vlek, R., Geuze, J., Nijholt, A., Ramsey, N., Haselager, P.,Vuurpijl, L., Gielen, S., and Desain, P. (2009). The brain-computer interface cycle. J. Neural Eng.,6(4):041001.

van Kesteren, M. T. R., Fernandez, G., Norris, D. G., and Hermans, E. J. (2010). Persistent schema-dependent hippocampal-neocortical connectivity during memory encoding and postencoding restin humans. Proc. Natl. Acad. Sci. U S A., 107(16):7550–7555.

van Strien, N. M., Cappaert, N. L. M., and Witter, M. P. (2009). The anatomy of memory: an interactiveoverview of the parahippocampal–hippocampal network. Nat. Rev. Neurosci., 10(4):272–282.

van Zwitten, M. V. A., Eumen, R. J. A., Grassmayr, T., and Bos, W. A. (1998). Symptoms of pseudobulbaraffect, anterograde amnesia and temporary cryptobiosis in brine shrimp Artemia NYOS followingprolonged alcohol exposure. Ann. Sci. Mon. Ult. Wee.

Vanderwolf, C. (1969). Hippocampal electrical activity and voluntary movement in the rat. Electroen-cephalogr. Clin. Neurophysiol., 26(4):407–418.

Varoquaux, G. and Thirion, B. (2014). How machine learning is shaping cognitive neuroimaging. Giga-science, 3(1).

Veen, B. V., Drongelen, W. V., Yuchtman, M., and Suzuki, A. (1997). Localization of brain electrical activityvia linearly constrainedminimum variance spatial filtering. IEEE Trans. Biomed. Eng., 44(9):867–880.

Vincent, J. L., Kahn, I., Snyder, A. Z., Raichle, M. E., and Buckner, R. L. (2008). Evidence for a frontoparietalcontrol system revealed by intrinsic functional connectivity. J. Neurophysiol., 100:3328–3342.

Vincent, J. L., Snyder, A. Z., Fox, M. D., Shannon, B. J., Andrews, J. R., Raichle, M. E., and Buckner, R. L.(2006). Coherent spontaneous activity identifies a hippocampal-parietal memory network. J. Neu-rophysiol., 96(6):3517–3531.

Vinogradova, O. (2001). Hippocampus as comparator: Role of the two input and two output systemsof the hippocampus in selection and registration of information. Hippocampus, 11(5):578–598.

Visser, R. M., Scholte, H. S., and Kindt, M. (2011). Associative learning increases trial-by-trial similarityof BOLD-MRI patterns. J. Neurosci., 31(33):12021–12028.

Wagner, A. D., Paré-Blagoev, E., Clark, J., and Poldrack, R. A. (2001). Recovering meaning. Neuron,31(2):329–338.

Wagner, A. D., Shannon, B. J., Kahn, I., and Buckner, R. L. (2005). Parietal lobe contributions to episodicmemory retrieval. Trends Cogn. Sci., 9(9):445–453.

Wallis, J. D. (2011). Cross-species studies of orbitofrontal cortex and value-based decision-making. Nat.Neurosci., 15(1):13–19.

Watrous, A. J., Tandon, N., Conner, C. R., Pieters, T., and Ekstrom, A. D. (2013). Frequency-specificnetwork connectivity increases underlie accurate spatiotemporal memory retrieval. Nat. Neurosci.,16:349–356.

Willenbockel, V. (2010). Controlling low-level image properties: the shine toolbox. Behav. Res.Methods,42:671–684.

Wills, T. J. (2005). Attractor dynamics in the hippocampal representation of the local environment.Science, 308(5723):873–876.

Winkler, A. M., Ridgway, G. R., Webster, M. A., Smith, S. M., and Nichols, T. E. (2014). Permutationinference for the general linear model. Neuroimage, 92:381–397.

Winson, J. (1978). Loss of hippocampal theta rhythm results in spatial memory deficit in the rat. Science,201(4351):160–163.

Wirth, S. (2003). Single neurons in themonkey hippocampus and learning of new associations. Science,300(5625):1578–1581.

Wirth, S., Avsar, E., Chiu, C. C., Sharma, V., Smith, A. C., Brown, E., and Suzuki, W. A. (2009). Trialoutcome and associative learning signals in the monkey hippocampus. Neuron, 61(6):930–940.

Wise, S. P. (2008). Forward frontal fields: phylogeny and fundamental function. Trends Neurosci.,31(12):599–608.

123

Wixted, J. T. (2014). Sparse and distributed coding of episodic memory in neurons of the human hip-pocampus. Proc. Natl Acad. Sci. USA, 111:9621–9626.

Wood, E. R., Dudchenko, P. A., and Eichenbaum, H. (1999). The global record of memory in hippocampalneuronal activity. Nature, 397:613–616.

Xu, W. and Südhof, T. C. (2013). A neural circuit for memory specificity and generalization. Science,339(6125):1290–1295.

Young, C. K. andMcNaughton, N. (2009). Coupling of theta oscillations between anterior and posteriormidline cortex and with the hippocampus in freely behaving rats. Cereb. Cortex, 19(1):24–40.

Zeithamova, D., Dominick, A. L., and Preston, A. R. (2012a). Hippocampal and ventral medial prefrontalactivation during retrieval-mediated learning supports novel inference. Neuron, 75:168–179.

Zeithamova, D. and Preston, A. R. (2010). Flexible memories: Differential roles for medial temporallobe and prefrontal cortex in cross-episode binding. J. Neurosci., 30(44):14676–14684.

Zeithamova, D., Schlichting, M. L., and Preston, A. R. (2012b). The hippocampus and inferential reason-ing: building memories to navigate future decisions. Front. Hum. Neurosci., 6.

Zhang, Y., Brady, M., and Smith, S. (2001). Segmentation of brain MR images through a hiddenmarkov randomfieldmodel and the expectation-maximization algorithm. IEEE Trans.Med. Imaging.,20(1):45–57.

Zheng, Z. Z. (2013). Multivoxel patterns reveal functionally differentiated networks underlying auditoryfeedback processing of speech. J. Neurosci., 33:4339–4348.

Nederlandse samenvatting

Mensen hebben een opmerkelijk vermogen om informatie te onthouden voor la-ter. We slaan herinneringen op, houden ze vast, combineren ze en we zijn in staatze weer op te halen. Deze vaardigheid is essentieel in ons dagelijks leven, bijvoor-beeld als je wilt onthouden of je iemand al eerder hebt ontmoet - misschien weltijdens dat ene speciale diner op een rondvaartboot. Uiteindelijk bepalen al onzeherinneringen samen wie wij zijn als persoon. Maar hoe maakt ons brein dit mo-gelijk? De precieze neurofysiologische processen die ten grondslag liggen aan hetcoderen, consolideren en het decoderen van herinneringen, zijn tot op heden onbe-kend. In dit proefschrift heb ik de hersenmechanismen van geheugen bestudeerdvanuit twee verschillende invalshoeken: hoe kan het brein herinneringen met elkaarin verband brengen en hoe verbinden deze herinneringen verschillende gespeciali-seerde hersengebieden? Deze processen heb ik onderzocht met behulp van cogni-tieve experimenten, moderne neuroimaging methoden, zoals functional MagneticResonance Imaging (fMRI), magnetoencephalography (MEG) en geavanceerde data-analyse technieken.

In Hoofdstuk 2 heb ik onderzocht waar in het brein de verschillende stukjes infor-matie van een gebeurtenis worden samengesmeed tot één coherente herinnering.Er zijn namelijk verschillende hersengebieden betrokken bij het verwerken van infor-matie over bijvoorbeeld de tijd en locatie van dat ene diner, de rondvaartboot, heteten en de personen. Mijn onderzoek laat zien dat de hippocampus - een gespecia-liseerde structuur diep in het brein die vanwege zijn vorm het ‘zeepaardje’ genoemdwordt - cruciaal is voor het combineren van deze losse elementen tot gedetailleerdeherinneringen. Hiervoor heb ik allereerst op basis van theoretische modellen tweecruciale eigenschappen vastgesteld, die in een hersengebied aanwezig zouden moe-ten zijn als het deze verbindingsfunctie zou hebben: het hersengebied moet 1) deinformatie bevatten van de herinnering en 2) in verbinding staan met veel anderehersengebieden, om zo de informatie te kunnen verzamelen en reactiveren. In deMRI-scanner gaf ik proefpersonen de opdracht om associaties te leren tussen plaat-jes van gezichten, huizen en lichamen. Vervolgens kregen ze één van de geassoci-eerde plaatjes te zien (bijvoorbeeld een gezicht) enmoesten ze zich herinneren welkplaatje daarbij hoorde (bijvoorbeeld een huis). Door de hersensignalen te analyse-ren, ontdekte ik dat de hippocampus informatie bevat over de gecombineerde herin-nering en tegelijkertijd intensief communiceert met andere hersengebieden tijdens

125

126

het herinneren van de geleerde associaties. Voor het eerst heb ik hiermee laten ziendat specifiek de hippocampus de twee cruciale eigenschappen heeft die nodig zijnvoor het combineren van informatie tot één coherente herinnering. De resultatenvan mijn onderzoek geven daarnaast inzicht in wat er gebeurt als de hippocampuswordt aangetast, zoals het geval is bij de ziekte van Alzheimer.

In Hoofdstuk 3 ben ik dieper ingegaan in het ontstaan van een associatieve herin-nering in de hippocampus. Hoe wordt deze herinnering gevormd en kunnen weeen verandering in het brein observeren met beeldvormende technieken? Uit eer-der onderzoek is gebleken dat de vorming van een nieuwe herinnering inderdaadmeetbare veranderingen in de hersenen teweegbrengt. Echter, is het daarbij ookmogelijk om als onderzoeker de herinneringen van een proefpersoon uit te lezen; tedecoderen uit de hersensignalen? Als dit decoderen mogelijk zou zijn, kunnen weiemands herinneringen in kaart brengen en continue monitoren hoe iemand ken-nis vergaart. In potentie kan dergelijke technologie van waarde zijn om educatie teverbeteren. Om deze open vraag te verkennen heb ik proefpersonen plaatjes vankleurrijke cirkels laten zien die in paren bij elkaar hoorden. Vervolgens heb ik gepro-beerd uit hersenscans te decoderen welke cirkels samen als paar door de proefper-soon waren onthouden, door de activiteit in de hersenen vóór en ná het leren metelkaar te vergelijken. Net als bij eerder onderzoek, observeerde ik een veranderingin de hippocampus in de groep proefpersonen: na het leren was de representatievan informatie in de hippocampus gemodificeerd. Vervolgens heb ik met verschil-lende algoritmes geprobeerd de herinneringen van individuen te decoderen uit dehersenscans. Met de huidige technieken bleek het echter nog niet mogelijk om ditop een betrouwbare manier te doen. De bevindingen van mijn onderzoek zijn des-alniettemin waardevol voor vervolgstudies. De experimentele opstelling, gebruiktemethoden en dataset met hersenscans kunnen dienen als een standaardtest vooronderzoek naar nieuwe algoritmes om herinneringen te decoderen, met daarbij dehuidige resultaten als benchmark. In de toekomst zal het daardoor met soortgelijkemethoden wellicht wel mogelijk worden om herinneringen uit te gaan lezen.

In Hoofdstuk 4 heb ik onderzocht hoe het brein in staat is om afzonderlijke herinne-ringen met elkaar te combineren. Het combineren van herinneringen is van belangvoor het leggen van associatieve verbanden tussen dingen die je nooit daadwerkelijksamen hebt gezien. Stel je ziet eenmoedermet een kind lopen. Wanneer je diezelfdemoeder even later met een ander kind ziet lopen, bedenk je zelf dat de twee kinde-ren waarschijnlijk broertjes of zusjes zijn, ook al heb je ze nooit samen gezien. Ietssoortgelijks gebeurt er bij een posttraumatische stressstoornis, bijvoorbeeld als ie-mand een schietincident heeft meegemaakt tijdens een oorlog: als diegene dan inhet normale leven een band hoort klappen, denkt hij direct weer aan de oorlog. Mijn

127

onderzoek laat zien wat er in de hersenen gebeurt zodra het schietincident met deklapband in verband wordt gebracht. Het antwoord: hersengolven. Ons brein staatnooit helemaal aan of uit. In werkelijkheid is activiteit van hersencellen netjes ge-organiseerd in golven die komen en gaan. Een bepaalde type trage hersengolven,de zogenaamde theta-golven, komt met name voor in de hippocampus. Uit eer-der onderzoek is gebleken dat theta-golven in de hippocampus een belangrijke rolspelen bij geheugen. Daarnaast is bekend dat de hippocampus intensief communi-ceert met de mediale prefrontale cortex. In Hoofdstuk 4 breng ik deze bevindingensamen, door te laten zien dat theta-golven het brein in staat stellen om te commu-niceren en zo herinneringen te combineren. Maar hoe meet je theta-golven? Datbleek niet gemakkelijk. In gezonde proefpersonen kun je hersengolven alleenmetenvan buitenaf, waardoor het lastig is om een signaal te krijgen vanuit de hippocam-pus, omdat deze structuur zeer diep in de hersenen verborgen zit. Daarom moestik geavanceerde wiskundige technieken gebruiken om de hersengolven te kunnenreconstrueren. Uit mijn onderzoek blijkt, dat iedere keer als iemand succesvol wasin het combineren van herinneringen, de sterkte van theta-golven in de hippocam-pus toeneemt. Bovendien bleek dat de theta-golven gekoppeld zijn aan de medialeprefrontaal cortex, een specifiek deel van de voorkwab betrokken bij het vergarenvan kennis. De gesynchroniseerde theta-golven zorgen er waarschijnlijk voor datver uit elkaar gelegen hersengebieden met elkaar kunnen communiceren: als dezetheta-golven er niet zijn, slagen de hersenen niet in het combineren van herinnerin-gen en het leggen van onderlinge verbanden. De resultaten van dit onderzoek zijnvan belang voor posttraumatische stressstoornis, omdat hier sprake lijkt te zijn vaneen ‘wildgroei’ aan negatieve associaties. Daarnaast zijn mijn bevindingen poten-tieel interessant voor de ontwikkeling van zelflerende computeralgoritmes, omdathet leggen van dit soort associatieve verbanden tot nu toe zeer lastig is voor com-puters. De resultaten vertellen ons ook wat over de werking van de hersenen in hetalgemeen. Het vermogen omherinneringen te combineren stelt ons in staat omwet-matigheden te ontdekken en op basis hiervan beslissingen te nemen. Uiteindelijk isdit hoe we kennis te vergaren over de wereld rondom ons en waarin we leven.

De experimentele hoofdstukken van dit proefschrift laten zien hoe het brein in staatis om bepaalde geheugenfuncties te vervullen. Hierbij is een bijzondere rol wegge-legd voor de hippocampus. De hippocampus combineert de verschillende aspectenvan de gebeurtenis (wie, wat, waar, wanneer) tot één coherente herinnering en istegelijkertijd sterk verbonden met vele andere hersengebieden. Na het aanmakenvan een herinnering, verandert de representatie van informatie in de hippocampus.Om verbanden te leggen tussen verschillende herinneringen, communiceert de hip-pocampus met de voorkwab via gesynchroniseerde theta-golven. Daarnaast zijn deontwikkelde methodes van belang voor toekomstige experimenten, bijvoorbeeld als

128

mediale pre-frontale cortex

voor: losse herinneringen

na:verbonden herinneringen

verbonden verbonden verbonden verbonden verbonden verbonden verbonden verbonden verbonden verbonden hersengebiedenhersengebiedenhersengebiedenhersengebiedenhersengebiedenhersengebiedenhersengebieden

verbonden hersengebieden

gebeurtenis

theta-golven

hippocampus

Hoe het brein herinneringen verbindt en hoe herinneringen het brein verbinden.Grafische samenvatting van de neurowetenschappelijke inzichten uit dit proefschrift. Geïsoleerdestukjes informatie (links) worden verbonden door een gebeurtenis (midden) en vormen één coherenteherinnering (rechts). Deze associaties zijn opgeslagen in de hippocampus die tijdens het ophalen vaneen herinnering sterk verbonden is met andere hersengebieden (pijlen). Theta-golven zorgen ervoordat de hippocampus kan communiceren met de mediale prefrontale cortex waardoor verbanden tus-sen oude en nieuwe herinneringen gelegd kunnen worden.

hersenonderzoekers herinneringen willen decoderen of theta-golven uit diepe her-senstructuren, zoals de hippocampus, willen meten. Afwijkingen in de synchronisa-tie van theta-golven, representaties of connectiviteit van de hippocampus zoudeneventueel als biomarker kunnen dienen voor diverse aandoeningen, zoals posttrau-matische stressstoornis of de ziekte van Alzheimer. Tot slot kunnen de inzichten enmethodieken uit de experimentele hoofdstukken gebruikt worden om educatie teverbeteren: door iteratief het geheugen en kennis van iemand te meten en met be-hulp van artificial intelligence (AI) te vergelijken met de kennis in een encyclopedie,zou een op maat gemaakt lesprogramma kunnen worden samengesteld. De bevin-dingen uit dit proefschrift vormen een aantal belangrijke puzzelstukjes, die ik hebtoegevoegd aan het neurowetenschappelijke onderzoeksveld van het menselijk ge-heugen. Deze puzzelstukjes helpen bij het leggen van de verbanden: hoe het breinherinneringen verbindt en hoe herinneringen het brein verbinden.

List of publications

Journal articles

1. Backus, A.R., Himmer, L., Doeller, C.F. (in preparation). Increased hippocam-pal neural pattern similarity of newly associated stimuli.

2. Bosch, S.E.*, Backus, A.R.*, and Doeller, C.F. (in preparation). Memory repre-sentations shift from hippocampus to medial frontal cortex through memoryconsolidation.†

3. Backus, A.R.*, Bosch, S.E.*, Ekman, M., Vicente-Grabovetsky, A., Doeller, C.F.(2016). Mnemonic convergence in the human hippocampus, Nature Commu-nications, 7(11991):1–9, doi: 10.1038/ncomms11991

4. Backus, A.R., Schoffelen, J-M., Szebényi, S., Hanslmayr, S., Doeller, C.F. (2016).Hippocampal-prefrontal theta oscillations support memory integration, Cur-rent Biology, 26:1–8, doi: 10.1016/j.cub.2015.12.048

5. van de Nieuwenhuijzen, M.E., Backus, A.R., Bahramisharif, A., Doeller, C.F.,Jensen, O., van Gerven, M.A. (2014). MEG-based decoding of the spatiotem-poral dynamics of visual category perception, NeuroImage, 83:1063-1073, doi:10.1016/j.neuroimage.2013.07.075

* denotes equal contributions.†Published as a chapter in the dissertation of Sander E. Bosch, (2016). Reactivating memories in hip-pocampus and neocortex, Chapter 5, 83–98, ISBN 978-94-6284-042-3

129

Acknowledgements

In December 2010, I was doing an internship project at the Donders Institute onhuman memory. As a student, I helped shape a small journal club: the hippocampusmeeting. Here, I met Christian Doeller, a young Principal Investigator setting up hisown group at the Institute. After he presented hisNature paper on human grid cellsduring one of the sessions, I was left impressed, but also had a lot of questions andideas. I decided to drop by his new office for a chat, and not long after, I joined hisnewly-founded research group to pursue a PhD. This event marked the beginning ofthe journey that resulted in this dissertation.

Christian, I can hardly express how grateful I am to you for everything. First of all,for giving me the opportunity to pursue a PhD in your research group. Secondly, foralways supporting me. In good times, for instance when we were brainstorming to-gether about exciting new experiments (not all of them have made it into this thesis,but we still learned a lot), and in bad times, when analysis results or editor decisionswere disappointing. Even when I announced I was leaving academia, you supportedmy decision. You were my main source of inspiration, with unbounded creativityand resourcefulness, that you used to solve every problem we encountered. I en-joyed and took example of your strong personality: your optimism, high ambition,perseverance, humor (April fools!), honesty, open-mindedness and, most of all, yourinexhaustible enthusiasm. Foremost, you taught me how to do research, but alsothe tips and tricks for success in the academic world. The skills I learned from youare of great use till this very day. Lastly, I want to emphasize that I am especiallygrateful for the freedom, responsibility and autonomy I enjoyed under your super-vision. Christian, my mentor, boss, friend and academic father, I have the utmostrespect for you. Thank you.

Every now and then you run into someone who is on the same wavelength. It is afeeling that your recognize from the very first encounter and you immediately knowthat this person is going the become a friend for life. Sander, you are one of thesepersons. Likeme, you have a keen eye for detail: letter case and coding style are veryimportant, documents must be versioned, and even the slightest graphics misalign-ment can keep you awake at night. But where I am sometimes a little bit too focusedon the task at hand, you are also a very people-oriented. This combination is prob-ably what makes us such a great team. I am very proud of our joint achievements,

131

132

such as the organization of the very first DoellerLab retreat (Schnitzel Samstag pro),our self-built introductory course, and last but not least, our joint fMRI project, re-sulting in both aNature Communications publication (Chapter 2) and a nice chapterin your thesis. I enjoyed your thoughtfulness, sociability and humor. Scanning ses-sions were never lonely or boring. You are always fun to hang out with - the LastManStanding at parties, be it in Café Vredenburg, the Guardian or playing billiards at DeStichtse. Although usually I have already left the party before or immediately afterthe so-called vodka phase, you champion the Sentimental Phase. I have many joyfulmemories of the trips we made, such as SFN social crashing or the philosophicalmidnight discussions backdropped by San Diago bay. Apart from the social aspect,I would like to add that you are also a very clever guy, with resourceful solutionsto various problems - the ideal sparring partner. You also possess a keen eye forpolitics and strategy, which we used to contrive cunning plans in room 1.18 together(harmless, as most of them were never executed, but fun). Maybe this treat is alsorelated to your significantly above-average win-rate for about every kind of game Iknow of... Sander, thank you for the work we did together as colleagues, and for allthe good times we had and will have as friends.

This brings me to Tobias, the third of the Doellerlab musketeers. You always offereda listening ear and thoughtful advice. Thank you for the great conversations onscience (and other things) and the comradeship during our time together. StarlightSalton Sea fish skeletons exploration with you and Sander is still one of the bestepisodic memories in my hippocampi.

The work presented in this thesis would not have been possible without the helpof two excellent students I was fortunate enough to be allowed to supervise. Szabi,your optimism and ability to performmultiple projects simultaneously never ceasedto amaze me. You always kept me sharp with your questions and quickly learned totackle problems on your own. I will never forget your Hungarian humor and good-bye goulash. Then secondly, Lea, I admire your cleverness, proactive attitude anduncanny time-management skills. I can still remember how you meticulously or-ganized the logistics of the experiments. Although the very first scan went wrong,you were in perfect control for all following. I actually thought it was more fun whenthings went wrong... Lea, Szabi, you were both great to work with and I am delightedby the fact that you both received excellent grades for the projects we did together.Thank you for making invaluable contributions to the experimental work presentedin Chapter 3 and Chapter 4. I wish you the best of luck in your careers.

In addition, I would like to thank my other collaborators for making the publicationsof this dissertation possible. Sasha, you brought RSA to the Donders and supported

133

me with your expert advice on fMRI analysis. Matthias, your network analyses andcreative ideas were vital to the success of Chapter 2. You also didme a favor by spark-ing my curiosity for the Python ecosystem. Jan-Mathijs, as the FieldTrip and MEGanalysis master, you assisted my quest to beam the hippocampus and taught meabout things like matrix algebra, lead fields, spatial filters and singular value decom-position. Marcel, thank you for introducing me to machine learning. I took note ofyour ever sharp and critical view on various cross-validation procedures, which hasproved invaluable for my role as data scientist. Ole, first let me thank you for seeingpotential in me as an MSc student and offering me an internship in your group atthe Donders, and secondly for introducing me to the wonderful world of neuronaloscillations. Simon, among other things, you helped me interpret the theta synchro-nization results for Chapter 4. Ali and Marieke, it was fun and educational to designand execute decoding experiments together, resulting in our nice NeuroImage pub-lication. Thanks to all of you.

To some extent, all DoellerLab members have contributed to this thesis, be it withtheir discussions, clever insights, comments, suggestions or just by providing anenormously fun and open work environment within the group. Most memorableare our SFN conference endeavors, including the infamous red Doellerlab wall, andthe joyous lab drinks. Thanks to Raphael, Ben, Sander, Toby, Sasha, Branka, Peter,Nils, Silvy, Jacob, Lorena, Lonsch, Nynke, Naomi, Stephanie, David, Staudi and allthe interns. Special mention Jacob: thanks for taking over my role as True Germanof the lab, by designing icons, structuring and organizing things, making sure RGB(or CMYK!) codes are all correct and everything is still shining either petrol green orDonders red (or pink or gold). And, Lonja, I enjoyed your sense of humor and thanksfor dragging me to the gym to counter the physical deterioration one suffers whenfinishing-up a PhD. Good luck to each and every one of you.

Besides the DoellerLab, I should thank all the other people at the Donders Institutefor making it such a great place to work. Summing up names here without forget-ting anyone is a futile exercise, so let me just say the following: the Donders is awonderful place to do research, with her superb neuroimaging experimentation re-sources, support staff, and a friendly atmosphere. The skewed amount of brainpowerof people surrounding you might make you feel humble, but pushes you to continu-ously learn and improve your skills. Thanks to the Donders Directors, TG, admin, allmy fellow PhD candidates, research assistants, interns, post-docs and research staff.Special mention Paul: thanks for teaching me how to operate the scanners and for(almost) always being up for cynical jokes.

In addition, I would like to spend some words on those who have been continu-

134

ously providing indirect support to make this dissertation possible. Firstly, I want tothank my parents, Chris and Roelf, and Oma Aukje, for their support and for being asignificant motivating force. I am assuming the (grand)son-with-PhD achievementwill make you very proud. Marjolein, thanks for helping me out with the final ef-forts leading up to the thesis defense event. In addition, thanks to you and Victor(bro’s!) for providing continuous sibling companionship (and rivalry). Thanks to theSeamonkeys, for saving me from solitary existence at high school, thereby enablingme to successfully complete the VWO and pursue an academic career. Thanks tomy homies at Warande 161, for providing the student living conditions needed tocomplete my studies. Thanks to my colleagues at BigData Republic, for recogniz-ing that the work presented in this thesis requires a skill set with great value outsideacademia. You offeredme an exciting outlook to the field of data science, while I wasfinishing this thesis. Last, but not least, thanks to my family, Marion and Josephine.DearMarion, thank you for taking care of Josephine (andme) during themany week-ends that were needed to complete this dissertation. Although it may not alwayshave seemed like it, you remind me of what is really important in life.

Curriculum Vitae

Alexander Rudolph Backus (1986) graduated from high school atthe Christelijk Lyceum Zeist in 2004, with a specialization in Na-ture & Health and ancient Greek language and culture. He thenpursued a BSc in Biomedical Sciences at Utrecht University, sup-plemented with courses in psychology, science communicationand a minor in philosophy. In 2011, Alexander obtained a MSc(cum laude) in Neuroscience & Cognition, with a specializationin Cognitive Neuroimaging, at Utrecht University.

During an internship in the group of dr. Dennis Schutter at theHelmholtz Institute, Utrecht University, Alexander investigatedthe neurophysiology of aggression, using electroencephalography and brain stimulationtechniques. In a second internship, Alexander studied human memory, using magnetoen-cephalography and machine learning techniques, in the groups of prof. dr. Ole Jensen anddr. Marcel van Gerven at the Donders Institute for Brain, Cognition and Behaviour, Rad-boud University. Subsequently, Alexander joined the newly-founded lab of prof. dr. Chris-tian Doeller, to pursue a PhD at the Donders Centre for Cognitive Neuroimaging. Here, hestudied the neural processes underlying memory, using advanced neuroimaging methods,such as functional magnetic resonance imaging. With his work published in journals suchas Current Biology and Nature Communications, Alexander made core contributions to theDoellerLab mission: cracking the code for memory.

In addition to his research activities, Alexander supported the research group and Institutein numerous ways, for instance by participating in various advisory committees, moderat-ing talks, designing graphics, leading the production of the Doellerlab website, structuringinternal processes and organizing team events. Together with colleague and friend SanderBosch, he developed, taught and coordinated a BSc-level course on cognitive neuroimaging.

After completing his PhD, Alexander joined data consulting firm BigData Republic, to workas a data scientist and big data consultant. Using core PhD skills, such as experiment design,advanced data analysis and data storytelling, he now provides strategic advice and hands-ondata science expertise to clients across various sectors, including retail, automotive, airline,finance and energy.

Alexander’s interests are in science and technology in general, and information processingsystems in particular: ranging from the human brain recalling its most precious memories,to artificial neural networks searching for meaningful patterns in big data.

135

Donders Graduate School for CognitiveNeuroscience

For a successful research Institute, it is vital to train the next generation of young scientists.To achieve this goal, theDonders Institute for Brain, Cognition andBehaviour established theDonders Graduate School for Cognitive Neuroscience (DGCN), which was officially recog-nised as a national graduate school in 2009. The Graduate School covers training at bothMaster’s and PhD level and provides an excellent educational context fully aligned with theresearch programme of the Donders Institute.

The school successfully attracts highly talented national and international students in biol-ogy, physics, psycholinguistics, psychology, behavioral science, medicine and related disci-plines. Selective admission and assessment centers guarantee the enrolment of the bestand most motivated students.

The DGCN tracks the career of PhD graduates carefully. More than 50% of PhD alumni showa continuation in academia with postdoc positions at top institutes worldwide, e.g. Stan-ford University, University of Oxford, University of Cambridge, UCL London, MPI Leipzig,HanyangUniversity in South Korea, NTNUNorway, University of Illinois, NorthWestern Uni-versity, Northeastern University in Boston, ETH Zürich, University of Vienna etc.. Positionsoutside academia spread among the following sectors: specialists in a medical environment,mainly in genetics, geriatrics, psychiatry and neurology. Specialists in a psychological envi-ronment, e.g. as specialist in neuropsychology, psychological diagnostics or therapy. Posi-tions in higher education as coordinators or lecturers. A smaller percentage enters businessas research consultants, analysts or head of research and development. Fewer graduatesstay in a research environment as lab coordinators, technical support or policy advisors. Up-coming possibilities are positions in the IT sector and management position in pharmaceu-tical industry. In general, the PhDs graduates almost invariably continue with high-qualitypositions that play an important role in our knowledge economy.

For more information on the DGCN as well as past and upcoming defenses please visit:http://www.ru.nl/donders/graduate-school/phd/

137

ISBN 978-94-6284119-2

295C

ON

NE

CT

ING

TH

E D

OT

S: T

HE

HIP

PO

CA

MPA

L M

EC

HA

NIS

MS

OF

AS

SO

CIA

TIV

E M

EM

OR

Y IN

HIP

PO

CA

MP

US

AN

D N

EO

CO

RT

EX

Alexander R. Backus